-//==-- SystemZISelDAGToDAG.cpp - A dag to dag inst selector for SystemZ ---===//
+//===-- SystemZISelDAGToDAG.cpp - A dag to dag inst selector for SystemZ --===//
//
// The LLVM Compiler Infrastructure
//
//
//===----------------------------------------------------------------------===//
-#include "SystemZ.h"
-#include "SystemZISelLowering.h"
#include "SystemZTargetMachine.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/Function.h"
-#include "llvm/Intrinsics.h"
-#include "llvm/CallingConv.h"
-#include "llvm/Constants.h"
-#include "llvm/CodeGen/MachineFrameInfo.h"
-#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
-#include "llvm/Target/TargetLowering.h"
-#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
using namespace llvm;
namespace {
- /// SystemZRRIAddressMode - This corresponds to rriaddr, but uses SDValue's
- /// instead of register numbers for the leaves of the matched tree.
- struct SystemZRRIAddressMode {
- enum {
- RegBase,
- FrameIndexBase
- } BaseType;
-
- struct { // This is really a union, discriminated by BaseType!
- SDValue Reg;
- int FrameIndex;
- } Base;
-
- SDValue IndexReg;
- int64_t Disp;
-
- SystemZRRIAddressMode()
- : BaseType(RegBase), IndexReg(), Disp(0) {
- }
+// Used to build addressing modes.
+struct SystemZAddressingMode {
+ // The shape of the address.
+ enum AddrForm {
+ // base+displacement
+ FormBD,
- void dump() {
- cerr << "SystemZRRIAddressMode " << this << '\n';
- if (BaseType == RegBase) {
- cerr << "Base.Reg ";
- if (Base.Reg.getNode() != 0) Base.Reg.getNode()->dump();
- else cerr << "nul";
- cerr << '\n';
- } else {
- cerr << " Base.FrameIndex " << Base.FrameIndex << '\n';
- }
- cerr << "IndexReg ";
- if (IndexReg.getNode() != 0) IndexReg.getNode()->dump();
- else cerr << "nul";
- cerr << " Disp " << Disp << '\n';
- }
+ // base+displacement+index for load and store operands
+ FormBDXNormal,
+
+ // base+displacement+index for load address operands
+ FormBDXLA,
+
+ // base+displacement+index+ADJDYNALLOC
+ FormBDXDynAlloc
};
-}
+ AddrForm Form;
+
+ // The type of displacement. The enum names here correspond directly
+ // to the definitions in SystemZOperand.td. We could split them into
+ // flags -- single/pair, 128-bit, etc. -- but it hardly seems worth it.
+ enum DispRange {
+ Disp12Only,
+ Disp12Pair,
+ Disp20Only,
+ Disp20Only128,
+ Disp20Pair
+ };
+ DispRange DR;
-/// SystemZDAGToDAGISel - SystemZ specific code to select SystemZ machine
-/// instructions for SelectionDAG operations.
-///
-namespace {
- class SystemZDAGToDAGISel : public SelectionDAGISel {
- SystemZTargetLowering &Lowering;
- const SystemZSubtarget &Subtarget;
+ // The parts of the address. The address is equivalent to:
+ //
+ // Base + Disp + Index + (IncludesDynAlloc ? ADJDYNALLOC : 0)
+ SDValue Base;
+ int64_t Disp;
+ SDValue Index;
+ bool IncludesDynAlloc;
- public:
- SystemZDAGToDAGISel(SystemZTargetMachine &TM, CodeGenOpt::Level OptLevel)
- : SelectionDAGISel(TM, OptLevel),
- Lowering(*TM.getTargetLowering()),
- Subtarget(*TM.getSubtargetImpl()) { }
+ SystemZAddressingMode(AddrForm form, DispRange dr)
+ : Form(form), DR(dr), Base(), Disp(0), Index(),
+ IncludesDynAlloc(false) {}
- virtual void InstructionSelect();
+ // True if the address can have an index register.
+ bool hasIndexField() { return Form != FormBD; }
- virtual const char *getPassName() const {
- return "SystemZ DAG->DAG Pattern Instruction Selection";
- }
+ // True if the address can (and must) include ADJDYNALLOC.
+ bool isDynAlloc() { return Form == FormBDXDynAlloc; }
- /// getI16Imm - Return a target constant with the specified value, of type
- /// i16.
- inline SDValue getI16Imm(uint64_t Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i16);
+ void dump() {
+ errs() << "SystemZAddressingMode " << this << '\n';
+
+ errs() << " Base ";
+ if (Base.getNode() != 0)
+ Base.getNode()->dump();
+ else
+ errs() << "null\n";
+
+ if (hasIndexField()) {
+ errs() << " Index ";
+ if (Index.getNode() != 0)
+ Index.getNode()->dump();
+ else
+ errs() << "null\n";
}
- /// getI32Imm - Return a target constant with the specified value, of type
- /// i32.
- inline SDValue getI32Imm(uint64_t Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i32);
+ errs() << " Disp " << Disp;
+ if (IncludesDynAlloc)
+ errs() << " + ADJDYNALLOC";
+ errs() << '\n';
+ }
+};
+
+// Return a mask with Count low bits set.
+static uint64_t allOnes(unsigned int Count) {
+ return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1;
+}
+
+// Represents operands 2 to 5 of the ROTATE AND ... SELECTED BITS operation
+// given by Opcode. The operands are: Input (R2), Start (I3), End (I4) and
+// Rotate (I5). The combined operand value is effectively:
+//
+// (or (rotl Input, Rotate), ~Mask)
+//
+// for RNSBG and:
+//
+// (and (rotl Input, Rotate), Mask)
+//
+// otherwise. The value has BitSize bits.
+struct RxSBGOperands {
+ RxSBGOperands(unsigned Op, SDValue N)
+ : Opcode(Op), BitSize(N.getValueType().getSizeInBits()),
+ Mask(allOnes(BitSize)), Input(N), Start(64 - BitSize), End(63),
+ Rotate(0) {}
+
+ unsigned Opcode;
+ unsigned BitSize;
+ uint64_t Mask;
+ SDValue Input;
+ unsigned Start;
+ unsigned End;
+ unsigned Rotate;
+};
+
+class SystemZDAGToDAGISel : public SelectionDAGISel {
+ const SystemZTargetLowering &Lowering;
+ const SystemZSubtarget &Subtarget;
+
+ // Used by SystemZOperands.td to create integer constants.
+ inline SDValue getImm(const SDNode *Node, uint64_t Imm) {
+ return CurDAG->getTargetConstant(Imm, Node->getValueType(0));
+ }
+
+ // Try to fold more of the base or index of AM into AM, where IsBase
+ // selects between the base and index.
+ bool expandAddress(SystemZAddressingMode &AM, bool IsBase);
+
+ // Try to describe N in AM, returning true on success.
+ bool selectAddress(SDValue N, SystemZAddressingMode &AM);
+
+ // Extract individual target operands from matched address AM.
+ void getAddressOperands(const SystemZAddressingMode &AM, EVT VT,
+ SDValue &Base, SDValue &Disp);
+ void getAddressOperands(const SystemZAddressingMode &AM, EVT VT,
+ SDValue &Base, SDValue &Disp, SDValue &Index);
+
+ // Try to match Addr as a FormBD address with displacement type DR.
+ // Return true on success, storing the base and displacement in
+ // Base and Disp respectively.
+ bool selectBDAddr(SystemZAddressingMode::DispRange DR, SDValue Addr,
+ SDValue &Base, SDValue &Disp);
+
+ // Try to match Addr as a FormBDX* address of form Form with
+ // displacement type DR. Return true on success, storing the base,
+ // displacement and index in Base, Disp and Index respectively.
+ bool selectBDXAddr(SystemZAddressingMode::AddrForm Form,
+ SystemZAddressingMode::DispRange DR, SDValue Addr,
+ SDValue &Base, SDValue &Disp, SDValue &Index);
+
+ // PC-relative address matching routines used by SystemZOperands.td.
+ bool selectPCRelAddress(SDValue Addr, SDValue &Target) {
+ if (Addr.getOpcode() == SystemZISD::PCREL_WRAPPER) {
+ Target = Addr.getOperand(0);
+ return true;
}
+ return false;
+ }
+
+ // BD matching routines used by SystemZOperands.td.
+ bool selectBDAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ return selectBDAddr(SystemZAddressingMode::Disp12Only, Addr, Base, Disp);
+ }
+ bool selectBDAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ return selectBDAddr(SystemZAddressingMode::Disp12Pair, Addr, Base, Disp);
+ }
+ bool selectBDAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ return selectBDAddr(SystemZAddressingMode::Disp20Only, Addr, Base, Disp);
+ }
+ bool selectBDAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ return selectBDAddr(SystemZAddressingMode::Disp20Pair, Addr, Base, Disp);
+ }
- // Include the pieces autogenerated from the target description.
- #include "SystemZGenDAGISel.inc"
+ // BDX matching routines used by SystemZOperands.td.
+ bool selectBDXAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
+ SystemZAddressingMode::Disp12Only,
+ Addr, Base, Disp, Index);
+ }
+ bool selectBDXAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
+ SystemZAddressingMode::Disp12Pair,
+ Addr, Base, Disp, Index);
+ }
+ bool selectDynAlloc12Only(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXDynAlloc,
+ SystemZAddressingMode::Disp12Only,
+ Addr, Base, Disp, Index);
+ }
+ bool selectBDXAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
+ SystemZAddressingMode::Disp20Only,
+ Addr, Base, Disp, Index);
+ }
+ bool selectBDXAddr20Only128(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
+ SystemZAddressingMode::Disp20Only128,
+ Addr, Base, Disp, Index);
+ }
+ bool selectBDXAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
+ SystemZAddressingMode::Disp20Pair,
+ Addr, Base, Disp, Index);
+ }
+ bool selectLAAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
+ SystemZAddressingMode::Disp12Pair,
+ Addr, Base, Disp, Index);
+ }
+ bool selectLAAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
+ SDValue &Index) {
+ return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
+ SystemZAddressingMode::Disp20Pair,
+ Addr, Base, Disp, Index);
+ }
- private:
- bool SelectAddrRRI(SDValue Op, SDValue Addr,
- SDValue &Base, SDValue &Index, SDValue &Disp);
- bool SelectLAAddr(SDValue Op, SDValue Addr,
- SDValue &Base, SDValue &Index, SDValue &Disp);
+ // Check whether (or Op (and X InsertMask)) is effectively an insertion
+ // of X into bits InsertMask of some Y != Op. Return true if so and
+ // set Op to that Y.
+ bool detectOrAndInsertion(SDValue &Op, uint64_t InsertMask);
+
+ // Try to fold some of RxSBG.Input into other fields of RxSBG.
+ // Return true on success.
+ bool expandRxSBG(RxSBGOperands &RxSBG);
+
+ // Return an undefined i64 value.
+ SDValue getUNDEF64(SDLoc DL);
+
+ // Convert N to VT, if it isn't already.
+ SDValue convertTo(SDLoc DL, EVT VT, SDValue N);
+
+ // Try to implement AND or shift node N using RISBG with the zero flag set.
+ // Return the selected node on success, otherwise return null.
+ SDNode *tryRISBGZero(SDNode *N);
+
+ // Try to use RISBG or Opcode to implement OR or XOR node N.
+ // Return the selected node on success, otherwise return null.
+ SDNode *tryRxSBG(SDNode *N, unsigned Opcode);
+
+ // If Op0 is null, then Node is a constant that can be loaded using:
+ //
+ // (Opcode UpperVal LowerVal)
+ //
+ // If Op0 is nonnull, then Node can be implemented using:
+ //
+ // (Opcode (Opcode Op0 UpperVal) LowerVal)
+ SDNode *splitLargeImmediate(unsigned Opcode, SDNode *Node, SDValue Op0,
+ uint64_t UpperVal, uint64_t LowerVal);
+
+ bool storeLoadCanUseMVC(SDNode *N) const;
+
+public:
+ SystemZDAGToDAGISel(SystemZTargetMachine &TM, CodeGenOpt::Level OptLevel)
+ : SelectionDAGISel(TM, OptLevel),
+ Lowering(*TM.getTargetLowering()),
+ Subtarget(*TM.getSubtargetImpl()) { }
+
+ // Override MachineFunctionPass.
+ virtual const char *getPassName() const LLVM_OVERRIDE {
+ return "SystemZ DAG->DAG Pattern Instruction Selection";
+ }
- SDNode *Select(SDValue Op);
- bool SelectAddrRI(const SDValue& Op, SDValue& Addr,
- SDValue &Base, SDValue &Disp);
- bool MatchAddress(SDValue N, SystemZRRIAddressMode &AM, unsigned Depth = 0);
- bool MatchAddressBase(SDValue N, SystemZRRIAddressMode &AM);
+ // Override SelectionDAGISel.
+ virtual SDNode *Select(SDNode *Node) LLVM_OVERRIDE;
+ virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
+ char ConstraintCode,
+ std::vector<SDValue> &OutOps)
+ LLVM_OVERRIDE;
- #ifndef NDEBUG
- unsigned Indent;
- #endif
- };
-} // end anonymous namespace
+ // Include the pieces autogenerated from the target description.
+ #include "SystemZGenDAGISel.inc"
+};
+} // end anonymous namespace
-/// createSystemZISelDag - This pass converts a legalized DAG into a
-/// SystemZ-specific DAG, ready for instruction scheduling.
-///
FunctionPass *llvm::createSystemZISelDag(SystemZTargetMachine &TM,
- CodeGenOpt::Level OptLevel) {
+ CodeGenOpt::Level OptLevel) {
return new SystemZDAGToDAGISel(TM, OptLevel);
}
-/// isImmSExt20 - This method tests to see if the node is either a 32-bit
-/// or 64-bit immediate, and if the value can be accurately represented as a
-/// sign extension from a 20-bit value. If so, this returns true and the
-/// immediate.
-static bool isImmSExt20(int64_t Val, int64_t &Imm) {
- if (Val >= -524288 && Val <= 524287) {
- Imm = Val;
+// Return true if Val should be selected as a displacement for an address
+// with range DR. Here we're interested in the range of both the instruction
+// described by DR and of any pairing instruction.
+static bool selectDisp(SystemZAddressingMode::DispRange DR, int64_t Val) {
+ switch (DR) {
+ case SystemZAddressingMode::Disp12Only:
+ return isUInt<12>(Val);
+
+ case SystemZAddressingMode::Disp12Pair:
+ case SystemZAddressingMode::Disp20Only:
+ case SystemZAddressingMode::Disp20Pair:
+ return isInt<20>(Val);
+
+ case SystemZAddressingMode::Disp20Only128:
+ return isInt<20>(Val) && isInt<20>(Val + 8);
+ }
+ llvm_unreachable("Unhandled displacement range");
+}
+
+// Change the base or index in AM to Value, where IsBase selects
+// between the base and index.
+static void changeComponent(SystemZAddressingMode &AM, bool IsBase,
+ SDValue Value) {
+ if (IsBase)
+ AM.Base = Value;
+ else
+ AM.Index = Value;
+}
+
+// The base or index of AM is equivalent to Value + ADJDYNALLOC,
+// where IsBase selects between the base and index. Try to fold the
+// ADJDYNALLOC into AM.
+static bool expandAdjDynAlloc(SystemZAddressingMode &AM, bool IsBase,
+ SDValue Value) {
+ if (AM.isDynAlloc() && !AM.IncludesDynAlloc) {
+ changeComponent(AM, IsBase, Value);
+ AM.IncludesDynAlloc = true;
return true;
}
return false;
}
-static bool isImmSExt20(SDNode *N, int64_t &Imm) {
- if (N->getOpcode() != ISD::Constant)
- return false;
+// The base of AM is equivalent to Base + Index. Try to use Index as
+// the index register.
+static bool expandIndex(SystemZAddressingMode &AM, SDValue Base,
+ SDValue Index) {
+ if (AM.hasIndexField() && !AM.Index.getNode()) {
+ AM.Base = Base;
+ AM.Index = Index;
+ return true;
+ }
+ return false;
+}
- return isImmSExt20(cast<ConstantSDNode>(N)->getSExtValue(), Imm);
+// The base or index of AM is equivalent to Op0 + Op1, where IsBase selects
+// between the base and index. Try to fold Op1 into AM's displacement.
+static bool expandDisp(SystemZAddressingMode &AM, bool IsBase,
+ SDValue Op0, ConstantSDNode *Op1) {
+ // First try adjusting the displacement.
+ int64_t TestDisp = AM.Disp + Op1->getSExtValue();
+ if (selectDisp(AM.DR, TestDisp)) {
+ changeComponent(AM, IsBase, Op0);
+ AM.Disp = TestDisp;
+ return true;
+ }
+
+ // We could consider forcing the displacement into a register and
+ // using it as an index, but it would need to be carefully tuned.
+ return false;
}
-static bool isImmSExt20(SDValue Op, int64_t &Imm) {
- return isImmSExt20(Op.getNode(), Imm);
+bool SystemZDAGToDAGISel::expandAddress(SystemZAddressingMode &AM,
+ bool IsBase) {
+ SDValue N = IsBase ? AM.Base : AM.Index;
+ unsigned Opcode = N.getOpcode();
+ if (Opcode == ISD::TRUNCATE) {
+ N = N.getOperand(0);
+ Opcode = N.getOpcode();
+ }
+ if (Opcode == ISD::ADD || CurDAG->isBaseWithConstantOffset(N)) {
+ SDValue Op0 = N.getOperand(0);
+ SDValue Op1 = N.getOperand(1);
+
+ unsigned Op0Code = Op0->getOpcode();
+ unsigned Op1Code = Op1->getOpcode();
+
+ if (Op0Code == SystemZISD::ADJDYNALLOC)
+ return expandAdjDynAlloc(AM, IsBase, Op1);
+ if (Op1Code == SystemZISD::ADJDYNALLOC)
+ return expandAdjDynAlloc(AM, IsBase, Op0);
+
+ if (Op0Code == ISD::Constant)
+ return expandDisp(AM, IsBase, Op1, cast<ConstantSDNode>(Op0));
+ if (Op1Code == ISD::Constant)
+ return expandDisp(AM, IsBase, Op0, cast<ConstantSDNode>(Op1));
+
+ if (IsBase && expandIndex(AM, Op0, Op1))
+ return true;
+ }
+ return false;
}
-/// Returns true if the address can be represented by a base register plus
-/// a signed 20-bit displacement [r+imm].
-bool SystemZDAGToDAGISel::SelectAddrRI(const SDValue& Op, SDValue& Addr,
- SDValue &Base, SDValue &Disp) {
- // FIXME dl should come from parent load or store, not from address
- DebugLoc dl = Addr.getDebugLoc();
- MVT VT = Addr.getValueType();
-
- if (Addr.getOpcode() == ISD::ADD) {
- int64_t Imm = 0;
- if (isImmSExt20(Addr.getOperand(1), Imm)) {
- Disp = CurDAG->getTargetConstant(Imm, MVT::i64);
- if (FrameIndexSDNode *FI =
- dyn_cast<FrameIndexSDNode>(Addr.getOperand(0))) {
- Base = CurDAG->getTargetFrameIndex(FI->getIndex(), VT);
- } else {
- Base = Addr.getOperand(0);
- }
- return true; // [r+i]
- }
- } else if (Addr.getOpcode() == ISD::OR) {
- int64_t Imm = 0;
- if (isImmSExt20(Addr.getOperand(1), Imm)) {
- // If this is an or of disjoint bitfields, we can codegen this as an add
- // (for better address arithmetic) if the LHS and RHS of the OR are
- // provably disjoint.
- APInt LHSKnownZero, LHSKnownOne;
- CurDAG->ComputeMaskedBits(Addr.getOperand(0),
- APInt::getAllOnesValue(Addr.getOperand(0)
- .getValueSizeInBits()),
- LHSKnownZero, LHSKnownOne);
-
- if ((LHSKnownZero.getZExtValue()|~(uint64_t)Imm) == ~0ULL) {
- // If all of the bits are known zero on the LHS or RHS, the add won't
- // carry.
- Base = Addr.getOperand(0);
- Disp = CurDAG->getTargetConstant(Imm, MVT::i64);
- return true;
- }
- }
- } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr)) {
- // Loading from a constant address.
-
- // If this address fits entirely in a 20-bit sext immediate field, codegen
- // this as "d(r0)"
- int64_t Imm;
- if (isImmSExt20(CN, Imm)) {
- Disp = CurDAG->getTargetConstant(Imm, MVT::i64);
- Base = CurDAG->getRegister(0, VT);
+// Return true if an instruction with displacement range DR should be
+// used for displacement value Val. selectDisp(DR, Val) must already hold.
+static bool isValidDisp(SystemZAddressingMode::DispRange DR, int64_t Val) {
+ assert(selectDisp(DR, Val) && "Invalid displacement");
+ switch (DR) {
+ case SystemZAddressingMode::Disp12Only:
+ case SystemZAddressingMode::Disp20Only:
+ case SystemZAddressingMode::Disp20Only128:
+ return true;
+
+ case SystemZAddressingMode::Disp12Pair:
+ // Use the other instruction if the displacement is too large.
+ return isUInt<12>(Val);
+
+ case SystemZAddressingMode::Disp20Pair:
+ // Use the other instruction if the displacement is small enough.
+ return !isUInt<12>(Val);
+ }
+ llvm_unreachable("Unhandled displacement range");
+}
+
+// Return true if Base + Disp + Index should be performed by LA(Y).
+static bool shouldUseLA(SDNode *Base, int64_t Disp, SDNode *Index) {
+ // Don't use LA(Y) for constants.
+ if (!Base)
+ return false;
+
+ // Always use LA(Y) for frame addresses, since we know that the destination
+ // register is almost always (perhaps always) going to be different from
+ // the frame register.
+ if (Base->getOpcode() == ISD::FrameIndex)
+ return true;
+
+ if (Disp) {
+ // Always use LA(Y) if there is a base, displacement and index.
+ if (Index)
return true;
- }
+
+ // Always use LA if the displacement is small enough. It should always
+ // be no worse than AGHI (and better if it avoids a move).
+ if (isUInt<12>(Disp))
+ return true;
+
+ // For similar reasons, always use LAY if the constant is too big for AGHI.
+ // LAY should be no worse than AGFI.
+ if (!isInt<16>(Disp))
+ return true;
+ } else {
+ // Don't use LA for plain registers.
+ if (!Index)
+ return false;
+
+ // Don't use LA for plain addition if the index operand is only used
+ // once. It should be a natural two-operand addition in that case.
+ if (Index->hasOneUse())
+ return false;
+
+ // Prefer addition if the second operation is sign-extended, in the
+ // hope of using AGF.
+ unsigned IndexOpcode = Index->getOpcode();
+ if (IndexOpcode == ISD::SIGN_EXTEND ||
+ IndexOpcode == ISD::SIGN_EXTEND_INREG)
+ return false;
}
- Disp = CurDAG->getTargetConstant(0, MVT::i64);
- if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Addr))
- Base = CurDAG->getTargetFrameIndex(FI->getIndex(), VT);
+ // Don't use LA for two-operand addition if either operand is only
+ // used once. The addition instructions are better in that case.
+ if (Base->hasOneUse())
+ return false;
+
+ return true;
+}
+
+// Return true if Addr is suitable for AM, updating AM if so.
+bool SystemZDAGToDAGISel::selectAddress(SDValue Addr,
+ SystemZAddressingMode &AM) {
+ // Start out assuming that the address will need to be loaded separately,
+ // then try to extend it as much as we can.
+ AM.Base = Addr;
+
+ // First try treating the address as a constant.
+ if (Addr.getOpcode() == ISD::Constant &&
+ expandDisp(AM, true, SDValue(), cast<ConstantSDNode>(Addr)))
+ ;
else
- Base = Addr;
- return true; // [r+0]
+ // Otherwise try expanding each component.
+ while (expandAddress(AM, true) ||
+ (AM.Index.getNode() && expandAddress(AM, false)))
+ continue;
+
+ // Reject cases where it isn't profitable to use LA(Y).
+ if (AM.Form == SystemZAddressingMode::FormBDXLA &&
+ !shouldUseLA(AM.Base.getNode(), AM.Disp, AM.Index.getNode()))
+ return false;
+
+ // Reject cases where the other instruction in a pair should be used.
+ if (!isValidDisp(AM.DR, AM.Disp))
+ return false;
+
+ // Make sure that ADJDYNALLOC is included where necessary.
+ if (AM.isDynAlloc() && !AM.IncludesDynAlloc)
+ return false;
+
+ DEBUG(AM.dump());
+ return true;
}
-/// MatchAddress - Add the specified node to the specified addressing mode,
-/// returning true if it cannot be done. This just pattern matches for the
-/// addressing mode.
-bool SystemZDAGToDAGISel::MatchAddress(SDValue N, SystemZRRIAddressMode &AM,
- unsigned Depth) {
- DebugLoc dl = N.getDebugLoc();
- DOUT << "MatchAddress: "; DEBUG(AM.dump());
- // Limit recursion.
- if (Depth > 5)
- return MatchAddressBase(N, AM);
-
- // FIXME: We can perform better here. If we have something like
- // (shift (add A, imm), N), we can try to reassociate stuff and fold shift of
- // imm into addressing mode.
- switch (N.getOpcode()) {
- default: break;
- case ISD::Constant: {
- int64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
- int64_t Imm;
- if (isImmSExt20(AM.Disp + Val, Imm)) {
- AM.Disp = Imm;
- return false;
- }
- break;
+// Insert a node into the DAG at least before Pos. This will reposition
+// the node as needed, and will assign it a node ID that is <= Pos's ID.
+// Note that this does *not* preserve the uniqueness of node IDs!
+// The selection DAG must no longer depend on their uniqueness when this
+// function is used.
+static void insertDAGNode(SelectionDAG *DAG, SDNode *Pos, SDValue N) {
+ if (N.getNode()->getNodeId() == -1 ||
+ N.getNode()->getNodeId() > Pos->getNodeId()) {
+ DAG->RepositionNode(Pos, N.getNode());
+ N.getNode()->setNodeId(Pos->getNodeId());
+ }
+}
+
+void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
+ EVT VT, SDValue &Base,
+ SDValue &Disp) {
+ Base = AM.Base;
+ if (!Base.getNode())
+ // Register 0 means "no base". This is mostly useful for shifts.
+ Base = CurDAG->getRegister(0, VT);
+ else if (Base.getOpcode() == ISD::FrameIndex) {
+ // Lower a FrameIndex to a TargetFrameIndex.
+ int64_t FrameIndex = cast<FrameIndexSDNode>(Base)->getIndex();
+ Base = CurDAG->getTargetFrameIndex(FrameIndex, VT);
+ } else if (Base.getValueType() != VT) {
+ // Truncate values from i64 to i32, for shifts.
+ assert(VT == MVT::i32 && Base.getValueType() == MVT::i64 &&
+ "Unexpected truncation");
+ SDLoc DL(Base);
+ SDValue Trunc = CurDAG->getNode(ISD::TRUNCATE, DL, VT, Base);
+ insertDAGNode(CurDAG, Base.getNode(), Trunc);
+ Base = Trunc;
}
- case ISD::FrameIndex:
- if (AM.BaseType == SystemZRRIAddressMode::RegBase
- && AM.Base.Reg.getNode() == 0) {
- AM.BaseType = SystemZRRIAddressMode::FrameIndexBase;
- AM.Base.FrameIndex = cast<FrameIndexSDNode>(N)->getIndex();
+ // Lower the displacement to a TargetConstant.
+ Disp = CurDAG->getTargetConstant(AM.Disp, VT);
+}
+
+void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
+ EVT VT, SDValue &Base,
+ SDValue &Disp, SDValue &Index) {
+ getAddressOperands(AM, VT, Base, Disp);
+
+ Index = AM.Index;
+ if (!Index.getNode())
+ // Register 0 means "no index".
+ Index = CurDAG->getRegister(0, VT);
+}
+
+bool SystemZDAGToDAGISel::selectBDAddr(SystemZAddressingMode::DispRange DR,
+ SDValue Addr, SDValue &Base,
+ SDValue &Disp) {
+ SystemZAddressingMode AM(SystemZAddressingMode::FormBD, DR);
+ if (!selectAddress(Addr, AM))
+ return false;
+
+ getAddressOperands(AM, Addr.getValueType(), Base, Disp);
+ return true;
+}
+
+bool SystemZDAGToDAGISel::selectBDXAddr(SystemZAddressingMode::AddrForm Form,
+ SystemZAddressingMode::DispRange DR,
+ SDValue Addr, SDValue &Base,
+ SDValue &Disp, SDValue &Index) {
+ SystemZAddressingMode AM(Form, DR);
+ if (!selectAddress(Addr, AM))
+ return false;
+
+ getAddressOperands(AM, Addr.getValueType(), Base, Disp, Index);
+ return true;
+}
+
+bool SystemZDAGToDAGISel::detectOrAndInsertion(SDValue &Op,
+ uint64_t InsertMask) {
+ // We're only interested in cases where the insertion is into some operand
+ // of Op, rather than into Op itself. The only useful case is an AND.
+ if (Op.getOpcode() != ISD::AND)
+ return false;
+
+ // We need a constant mask.
+ ConstantSDNode *MaskNode =
+ dyn_cast<ConstantSDNode>(Op.getOperand(1).getNode());
+ if (!MaskNode)
+ return false;
+
+ // It's not an insertion of Op.getOperand(0) if the two masks overlap.
+ uint64_t AndMask = MaskNode->getZExtValue();
+ if (InsertMask & AndMask)
+ return false;
+
+ // It's only an insertion if all bits are covered or are known to be zero.
+ // The inner check covers all cases but is more expensive.
+ uint64_t Used = allOnes(Op.getValueType().getSizeInBits());
+ if (Used != (AndMask | InsertMask)) {
+ APInt KnownZero, KnownOne;
+ CurDAG->ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne);
+ if (Used != (AndMask | InsertMask | KnownZero.getZExtValue()))
return false;
- }
- break;
+ }
- case ISD::SUB: {
- // Given A-B, if A can be completely folded into the address and
- // the index field with the index field unused, use -B as the index.
- // This is a win if a has multiple parts that can be folded into
- // the address. Also, this saves a mov if the base register has
- // other uses, since it avoids a two-address sub instruction, however
- // it costs an additional mov if the index register has other uses.
-
- // Test if the LHS of the sub can be folded.
- SystemZRRIAddressMode Backup = AM;
- if (MatchAddress(N.getNode()->getOperand(0), AM, Depth+1)) {
- AM = Backup;
- break;
- }
- // Test if the index field is free for use.
- if (AM.IndexReg.getNode()) {
- AM = Backup;
- break;
- }
+ Op = Op.getOperand(0);
+ return true;
+}
- // If the base is a register with multiple uses, this transformation may
- // save a mov. Otherwise it's probably better not to do it.
- if (AM.BaseType == SystemZRRIAddressMode::RegBase &&
- (!AM.Base.Reg.getNode() || AM.Base.Reg.getNode()->hasOneUse())) {
- AM = Backup;
- break;
- }
+// Return true if Mask matches the regexp 0*1+0*, given that zero masks
+// have already been filtered out. Store the first set bit in LSB and
+// the number of set bits in Length if so.
+static bool isStringOfOnes(uint64_t Mask, unsigned &LSB, unsigned &Length) {
+ unsigned First = findFirstSet(Mask);
+ uint64_t Top = (Mask >> First) + 1;
+ if ((Top & -Top) == Top) {
+ LSB = First;
+ Length = findFirstSet(Top);
+ return true;
+ }
+ return false;
+}
- // Ok, the transformation is legal and appears profitable. Go for it.
- SDValue RHS = N.getNode()->getOperand(1);
- SDValue Zero = CurDAG->getConstant(0, N.getValueType());
- SDValue Neg = CurDAG->getNode(ISD::SUB, dl, N.getValueType(), Zero, RHS);
- AM.IndexReg = Neg;
-
- // Insert the new nodes into the topological ordering.
- if (Zero.getNode()->getNodeId() == -1 ||
- Zero.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), Zero.getNode());
- Zero.getNode()->setNodeId(N.getNode()->getNodeId());
- }
- if (Neg.getNode()->getNodeId() == -1 ||
- Neg.getNode()->getNodeId() > N.getNode()->getNodeId()) {
- CurDAG->RepositionNode(N.getNode(), Neg.getNode());
- Neg.getNode()->setNodeId(N.getNode()->getNodeId());
- }
+// Try to update RxSBG so that only the bits of RxSBG.Input in Mask are used.
+// Return true on success.
+static bool refineRxSBGMask(RxSBGOperands &RxSBG, uint64_t Mask) {
+ if (RxSBG.Rotate != 0)
+ Mask = (Mask << RxSBG.Rotate) | (Mask >> (64 - RxSBG.Rotate));
+ Mask &= RxSBG.Mask;
+
+ // Reject trivial all-zero masks.
+ if (Mask == 0)
return false;
+
+ // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of
+ // the msb and End specifies the index of the lsb.
+ unsigned LSB, Length;
+ if (isStringOfOnes(Mask, LSB, Length)) {
+ RxSBG.Mask = Mask;
+ RxSBG.Start = 63 - (LSB + Length - 1);
+ RxSBG.End = 63 - LSB;
+ return true;
}
- case ISD::ADD: {
- SystemZRRIAddressMode Backup = AM;
- if (!MatchAddress(N.getNode()->getOperand(0), AM, Depth+1) &&
- !MatchAddress(N.getNode()->getOperand(1), AM, Depth+1))
- return false;
- AM = Backup;
- if (!MatchAddress(N.getNode()->getOperand(1), AM, Depth+1) &&
- !MatchAddress(N.getNode()->getOperand(0), AM, Depth+1))
+ // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb
+ // of the low 1s and End specifies the lsb of the high 1s.
+ if (isStringOfOnes(Mask ^ allOnes(RxSBG.BitSize), LSB, Length)) {
+ assert(LSB > 0 && "Bottom bit must be set");
+ assert(LSB + Length < RxSBG.BitSize && "Top bit must be set");
+ RxSBG.Mask = Mask;
+ RxSBG.Start = 63 - (LSB - 1);
+ RxSBG.End = 63 - (LSB + Length);
+ return true;
+ }
+
+ return false;
+}
+
+// RxSBG.Input is a shift of Count bits in the direction given by IsLeft.
+// Return true if the result depends on the signs or zeros that are
+// shifted in.
+static bool shiftedInBitsMatter(RxSBGOperands &RxSBG, uint64_t Count,
+ bool IsLeft) {
+ // Work out which bits of the shift result are zeros or sign copies.
+ uint64_t ShiftedIn = allOnes(Count);
+ if (!IsLeft)
+ ShiftedIn <<= RxSBG.BitSize - Count;
+
+ // Rotate that mask in the same way as RxSBG.Input is rotated.
+ if (RxSBG.Rotate != 0)
+ ShiftedIn = ((ShiftedIn << RxSBG.Rotate) |
+ (ShiftedIn >> (64 - RxSBG.Rotate)));
+
+ // Fail if any of the zero or sign bits are used.
+ return (ShiftedIn & RxSBG.Mask) != 0;
+}
+
+bool SystemZDAGToDAGISel::expandRxSBG(RxSBGOperands &RxSBG) {
+ SDValue N = RxSBG.Input;
+ unsigned Opcode = N.getOpcode();
+ switch (Opcode) {
+ case ISD::AND: {
+ if (RxSBG.Opcode == SystemZ::RNSBG)
return false;
- AM = Backup;
-
- // If we couldn't fold both operands into the address at the same time,
- // see if we can just put each operand into a register and fold at least
- // the add.
- if (AM.BaseType == SystemZRRIAddressMode::RegBase &&
- !AM.Base.Reg.getNode() && !AM.IndexReg.getNode()) {
- AM.Base.Reg = N.getNode()->getOperand(0);
- AM.IndexReg = N.getNode()->getOperand(1);
+
+ ConstantSDNode *MaskNode =
+ dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ if (!MaskNode)
return false;
+
+ SDValue Input = N.getOperand(0);
+ uint64_t Mask = MaskNode->getZExtValue();
+ if (!refineRxSBGMask(RxSBG, Mask)) {
+ // If some bits of Input are already known zeros, those bits will have
+ // been removed from the mask. See if adding them back in makes the
+ // mask suitable.
+ APInt KnownZero, KnownOne;
+ CurDAG->ComputeMaskedBits(Input, KnownZero, KnownOne);
+ Mask |= KnownZero.getZExtValue();
+ if (!refineRxSBGMask(RxSBG, Mask))
+ return false;
}
- break;
+ RxSBG.Input = Input;
+ return true;
}
- case ISD::OR:
- // Handle "X | C" as "X + C" iff X is known to have C bits clear.
- if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
- SystemZRRIAddressMode Backup = AM;
- int64_t Offset = CN->getSExtValue();
- int64_t Imm;
- // Start with the LHS as an addr mode.
- if (!MatchAddress(N.getOperand(0), AM, Depth+1) &&
- // The resultant disp must fit in 20-bits.
- isImmSExt20(AM.Disp + Offset, Imm) &&
- // Check to see if the LHS & C is zero.
- CurDAG->MaskedValueIsZero(N.getOperand(0), CN->getAPIntValue())) {
- AM.Disp = Imm;
+ case ISD::OR: {
+ if (RxSBG.Opcode != SystemZ::RNSBG)
+ return false;
+
+ ConstantSDNode *MaskNode =
+ dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ if (!MaskNode)
+ return false;
+
+ SDValue Input = N.getOperand(0);
+ uint64_t Mask = ~MaskNode->getZExtValue();
+ if (!refineRxSBGMask(RxSBG, Mask)) {
+ // If some bits of Input are already known ones, those bits will have
+ // been removed from the mask. See if adding them back in makes the
+ // mask suitable.
+ APInt KnownZero, KnownOne;
+ CurDAG->ComputeMaskedBits(Input, KnownZero, KnownOne);
+ Mask &= ~KnownOne.getZExtValue();
+ if (!refineRxSBGMask(RxSBG, Mask))
return false;
- }
- AM = Backup;
}
- break;
+ RxSBG.Input = Input;
+ return true;
}
- return MatchAddressBase(N, AM);
-}
+ case ISD::ROTL: {
+ // Any 64-bit rotate left can be merged into the RxSBG.
+ if (RxSBG.BitSize != 64)
+ return false;
+ ConstantSDNode *CountNode
+ = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ if (!CountNode)
+ return false;
+
+ RxSBG.Rotate = (RxSBG.Rotate + CountNode->getZExtValue()) & 63;
+ RxSBG.Input = N.getOperand(0);
+ return true;
+ }
+
+ case ISD::SHL: {
+ ConstantSDNode *CountNode =
+ dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ if (!CountNode)
+ return false;
-/// MatchAddressBase - Helper for MatchAddress. Add the specified node to the
-/// specified addressing mode without any further recursion.
-bool SystemZDAGToDAGISel::MatchAddressBase(SDValue N,
- SystemZRRIAddressMode &AM) {
- // Is the base register already occupied?
- if (AM.BaseType != SystemZRRIAddressMode::RegBase || AM.Base.Reg.getNode()) {
- // If so, check to see if the scale index register is set.
- if (AM.IndexReg.getNode() == 0) {
- AM.IndexReg = N;
+ uint64_t Count = CountNode->getZExtValue();
+ if (Count < 1 || Count >= RxSBG.BitSize)
return false;
+
+ if (RxSBG.Opcode == SystemZ::RNSBG) {
+ // Treat (shl X, count) as (rotl X, size-count) as long as the bottom
+ // count bits from RxSBG.Input are ignored.
+ if (shiftedInBitsMatter(RxSBG, Count, true))
+ return false;
+ } else {
+ // Treat (shl X, count) as (and (rotl X, count), ~0<<count).
+ if (!refineRxSBGMask(RxSBG, allOnes(RxSBG.BitSize - Count) << Count))
+ return false;
}
- // Otherwise, we cannot select it.
+ RxSBG.Rotate = (RxSBG.Rotate + Count) & 63;
+ RxSBG.Input = N.getOperand(0);
return true;
}
- // Default, generate it as a register.
- AM.BaseType = SystemZRRIAddressMode::RegBase;
- AM.Base.Reg = N;
- return false;
-}
+ case ISD::SRL:
+ case ISD::SRA: {
+ ConstantSDNode *CountNode =
+ dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ if (!CountNode)
+ return false;
-/// Returns true if the address can be represented by a base register plus
-/// index register plus a signed 20-bit displacement [base + idx + imm].
-bool SystemZDAGToDAGISel::SelectAddrRRI(SDValue Op, SDValue Addr,
- SDValue &Base, SDValue &Index, SDValue &Disp) {
- SystemZRRIAddressMode AM;
- bool Done = false;
-
- if (!Addr.hasOneUse()) {
- unsigned Opcode = Addr.getOpcode();
- if (Opcode != ISD::Constant && Opcode != ISD::FrameIndex) {
- // If we are able to fold N into addressing mode, then we'll allow it even
- // if N has multiple uses. In general, addressing computation is used as
- // addresses by all of its uses. But watch out for CopyToReg uses, that
- // means the address computation is liveout. It will be computed by a LA
- // so we want to avoid computing the address twice.
- for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
- UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
- if (UI->getOpcode() == ISD::CopyToReg) {
- MatchAddressBase(Addr, AM);
- Done = true;
- break;
- }
- }
+ uint64_t Count = CountNode->getZExtValue();
+ if (Count < 1 || Count >= RxSBG.BitSize)
+ return false;
+
+ if (RxSBG.Opcode == SystemZ::RNSBG || Opcode == ISD::SRA) {
+ // Treat (srl|sra X, count) as (rotl X, size-count) as long as the top
+ // count bits from RxSBG.Input are ignored.
+ if (shiftedInBitsMatter(RxSBG, Count, false))
+ return false;
+ } else {
+ // Treat (srl X, count), mask) as (and (rotl X, size-count), ~0>>count),
+ // which is similar to SLL above.
+ if (!refineRxSBGMask(RxSBG, allOnes(RxSBG.BitSize - Count)))
+ return false;
}
+
+ RxSBG.Rotate = (RxSBG.Rotate - Count) & 63;
+ RxSBG.Input = N.getOperand(0);
+ return true;
}
- if (!Done && MatchAddress(Addr, AM))
+ default:
return false;
+ }
+}
- DOUT << "MatchAddress (final): "; DEBUG(AM.dump());
+SDValue SystemZDAGToDAGISel::getUNDEF64(SDLoc DL) {
+ SDNode *N = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::i64);
+ return SDValue(N, 0);
+}
- MVT VT = Addr.getValueType();
- if (AM.BaseType == SystemZRRIAddressMode::RegBase) {
- if (!AM.Base.Reg.getNode())
- AM.Base.Reg = CurDAG->getRegister(0, VT);
+SDValue SystemZDAGToDAGISel::convertTo(SDLoc DL, EVT VT, SDValue N) {
+ if (N.getValueType() == MVT::i32 && VT == MVT::i64) {
+ SDValue Index = CurDAG->getTargetConstant(SystemZ::subreg_32bit, MVT::i64);
+ SDNode *Insert = CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG,
+ DL, VT, getUNDEF64(DL), N, Index);
+ return SDValue(Insert, 0);
+ }
+ if (N.getValueType() == MVT::i64 && VT == MVT::i32) {
+ SDValue Index = CurDAG->getTargetConstant(SystemZ::subreg_32bit, MVT::i64);
+ SDNode *Extract = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
+ DL, VT, N, Index);
+ return SDValue(Extract, 0);
}
+ assert(N.getValueType() == VT && "Unexpected value types");
+ return N;
+}
- if (!AM.IndexReg.getNode())
- AM.IndexReg = CurDAG->getRegister(0, VT);
+SDNode *SystemZDAGToDAGISel::tryRISBGZero(SDNode *N) {
+ RxSBGOperands RISBG(SystemZ::RISBG, SDValue(N, 0));
+ unsigned Count = 0;
+ while (expandRxSBG(RISBG))
+ Count += 1;
+ // Prefer to use normal shift instructions over RISBG, since they can handle
+ // all cases and are sometimes shorter. Prefer to use RISBG for ANDs though,
+ // since it is effectively a three-operand instruction in this case,
+ // and since it can handle some masks that AND IMMEDIATE can't.
+ if (Count < (N->getOpcode() == ISD::AND ? 1U : 2U))
+ return 0;
+
+ // Prefer register extensions like LLC over RISBG.
+ if (RISBG.Rotate == 0 &&
+ (RISBG.Start == 32 || RISBG.Start == 48 || RISBG.Start == 56) &&
+ RISBG.End == 63)
+ return 0;
+
+ EVT VT = N->getValueType(0);
+ SDValue Ops[5] = {
+ getUNDEF64(SDLoc(N)),
+ convertTo(SDLoc(N), MVT::i64, RISBG.Input),
+ CurDAG->getTargetConstant(RISBG.Start, MVT::i32),
+ CurDAG->getTargetConstant(RISBG.End | 128, MVT::i32),
+ CurDAG->getTargetConstant(RISBG.Rotate, MVT::i32)
+ };
+ N = CurDAG->getMachineNode(SystemZ::RISBG, SDLoc(N), MVT::i64, Ops);
+ return convertTo(SDLoc(N), VT, SDValue(N, 0)).getNode();
+}
- if (AM.BaseType == SystemZRRIAddressMode::RegBase)
- Base = AM.Base.Reg;
- else
- Base = CurDAG->getTargetFrameIndex(AM.Base.FrameIndex, TLI.getPointerTy());
- Index = AM.IndexReg;
- Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i64);
+SDNode *SystemZDAGToDAGISel::tryRxSBG(SDNode *N, unsigned Opcode) {
+ // Try treating each operand of N as the second operand of the RxSBG
+ // and see which goes deepest.
+ RxSBGOperands RxSBG[] = {
+ RxSBGOperands(Opcode, N->getOperand(0)),
+ RxSBGOperands(Opcode, N->getOperand(1))
+ };
+ unsigned Count[] = { 0, 0 };
+ for (unsigned I = 0; I < 2; ++I)
+ while (expandRxSBG(RxSBG[I]))
+ Count[I] += 1;
+
+ // Do nothing if neither operand is suitable.
+ if (Count[0] == 0 && Count[1] == 0)
+ return 0;
+
+ // Pick the deepest second operand.
+ unsigned I = Count[0] > Count[1] ? 0 : 1;
+ SDValue Op0 = N->getOperand(I ^ 1);
+
+ // Prefer IC for character insertions from memory.
+ if (Opcode == SystemZ::ROSBG && (RxSBG[I].Mask & 0xff) == 0)
+ if (LoadSDNode *Load = dyn_cast<LoadSDNode>(Op0.getNode()))
+ if (Load->getMemoryVT() == MVT::i8)
+ return 0;
+
+ // See whether we can avoid an AND in the first operand by converting
+ // ROSBG to RISBG.
+ if (Opcode == SystemZ::ROSBG && detectOrAndInsertion(Op0, RxSBG[I].Mask))
+ Opcode = SystemZ::RISBG;
+
+ EVT VT = N->getValueType(0);
+ SDValue Ops[5] = {
+ convertTo(SDLoc(N), MVT::i64, Op0),
+ convertTo(SDLoc(N), MVT::i64, RxSBG[I].Input),
+ CurDAG->getTargetConstant(RxSBG[I].Start, MVT::i32),
+ CurDAG->getTargetConstant(RxSBG[I].End, MVT::i32),
+ CurDAG->getTargetConstant(RxSBG[I].Rotate, MVT::i32)
+ };
+ N = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i64, Ops);
+ return convertTo(SDLoc(N), VT, SDValue(N, 0)).getNode();
+}
- return true;
+SDNode *SystemZDAGToDAGISel::splitLargeImmediate(unsigned Opcode, SDNode *Node,
+ SDValue Op0, uint64_t UpperVal,
+ uint64_t LowerVal) {
+ EVT VT = Node->getValueType(0);
+ SDLoc DL(Node);
+ SDValue Upper = CurDAG->getConstant(UpperVal, VT);
+ if (Op0.getNode())
+ Upper = CurDAG->getNode(Opcode, DL, VT, Op0, Upper);
+ Upper = SDValue(Select(Upper.getNode()), 0);
+
+ SDValue Lower = CurDAG->getConstant(LowerVal, VT);
+ SDValue Or = CurDAG->getNode(Opcode, DL, VT, Upper, Lower);
+ return Or.getNode();
}
-/// SelectLAAddr - it calls SelectAddr and determines if the maximal addressing
-/// mode it matches can be cost effectively emitted as an LA/LAY instruction.
-bool SystemZDAGToDAGISel::SelectLAAddr(SDValue Op, SDValue Addr,
- SDValue &Base, SDValue &Index, SDValue &Disp) {
- SystemZRRIAddressMode AM;
+// N is a (store (load ...), ...) pattern. Return true if it can use MVC.
+bool SystemZDAGToDAGISel::storeLoadCanUseMVC(SDNode *N) const {
+ StoreSDNode *Store = cast<StoreSDNode>(N);
+ LoadSDNode *Load = cast<LoadSDNode>(Store->getValue().getNode());
- if (MatchAddress(Addr, AM))
+ // MVC is logically a bytewise copy, so can't be used for volatile accesses.
+ if (Load->isVolatile() || Store->isVolatile())
return false;
- MVT VT = Addr.getValueType();
- unsigned Complexity = 0;
- if (AM.BaseType == SystemZRRIAddressMode::RegBase)
- if (AM.Base.Reg.getNode())
- Complexity = 1;
- else
- AM.Base.Reg = CurDAG->getRegister(0, VT);
- else if (AM.BaseType == SystemZRRIAddressMode::FrameIndexBase)
- Complexity = 4;
-
- if (AM.IndexReg.getNode())
- Complexity += 1;
- else
- AM.IndexReg = CurDAG->getRegister(0, VT);
+ // Prefer not to use MVC if either address can use ... RELATIVE LONG
+ // instructions.
+ assert(Load->getMemoryVT() == Store->getMemoryVT() &&
+ "Should already have checked that the types match");
+ uint64_t Size = Load->getMemoryVT().getStoreSize();
+ if (Size > 1 && Size <= 8) {
+ // Prefer LHRL, LRL and LGRL.
+ if (Load->getBasePtr().getOpcode() == SystemZISD::PCREL_WRAPPER)
+ return false;
+ // Prefer STHRL, STRL and STGRL.
+ if (Store->getBasePtr().getOpcode() == SystemZISD::PCREL_WRAPPER)
+ return false;
+ }
- if (AM.Disp && (AM.Base.Reg.getNode() || AM.IndexReg.getNode()))
- Complexity += 1;
+ // There's no chance of overlap if the load is invariant.
+ if (Load->isInvariant())
+ return true;
- if (Complexity > 2) {
- if (AM.BaseType == SystemZRRIAddressMode::RegBase)
- Base = AM.Base.Reg;
- else
- Base = CurDAG->getTargetFrameIndex(AM.Base.FrameIndex,
- TLI.getPointerTy());
- Index = AM.IndexReg;
- Disp = CurDAG->getTargetConstant(AM.Disp, MVT::i64);
+ // If both operands are aligned, they must be equal or not overlap.
+ if (Load->getAlignment() >= Size && Store->getAlignment() >= Size)
return true;
- }
- return false;
-}
+ // Otherwise we need to check whether there's an alias.
+ const Value *V1 = Load->getSrcValue();
+ const Value *V2 = Store->getSrcValue();
+ if (!V1 || !V2)
+ return false;
-/// InstructionSelect - This callback is invoked by
-/// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
-void SystemZDAGToDAGISel::InstructionSelect() {
- DEBUG(BB->dump());
-
- // Codegen the basic block.
-#ifndef NDEBUG
- DOUT << "===== Instruction selection begins:\n";
- Indent = 0;
-#endif
- SelectRoot(*CurDAG);
-#ifndef NDEBUG
- DOUT << "===== Instruction selection ends:\n";
-#endif
-
- CurDAG->RemoveDeadNodes();
+ int64_t End1 = Load->getSrcValueOffset() + Size;
+ int64_t End2 = Store->getSrcValueOffset() + Size;
+ return !AA->alias(AliasAnalysis::Location(V1, End1, Load->getTBAAInfo()),
+ AliasAnalysis::Location(V2, End2, Store->getTBAAInfo()));
}
-SDNode *SystemZDAGToDAGISel::Select(SDValue Op) {
- SDNode *Node = Op.getNode();
- DebugLoc dl = Op.getDebugLoc();
-
+SDNode *SystemZDAGToDAGISel::Select(SDNode *Node) {
// Dump information about the Node being selected
- #ifndef NDEBUG
- DOUT << std::string(Indent, ' ') << "Selecting: ";
- DEBUG(Node->dump(CurDAG));
- DOUT << "\n";
- Indent += 2;
- #endif
+ DEBUG(errs() << "Selecting: "; Node->dump(CurDAG); errs() << "\n");
// If we have a custom node, we already have selected!
if (Node->isMachineOpcode()) {
- #ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "== ";
- DEBUG(Node->dump(CurDAG));
- DOUT << "\n";
- Indent -= 2;
- #endif
- return NULL;
+ DEBUG(errs() << "== "; Node->dump(CurDAG); errs() << "\n");
+ return 0;
}
- // Select the default instruction
- SDNode *ResNode = SelectCode(Op);
+ unsigned Opcode = Node->getOpcode();
+ SDNode *ResNode = 0;
+ switch (Opcode) {
+ case ISD::OR:
+ if (Node->getOperand(1).getOpcode() != ISD::Constant)
+ ResNode = tryRxSBG(Node, SystemZ::ROSBG);
+ goto or_xor;
+
+ case ISD::XOR:
+ if (Node->getOperand(1).getOpcode() != ISD::Constant)
+ ResNode = tryRxSBG(Node, SystemZ::RXSBG);
+ // Fall through.
+ or_xor:
+ // If this is a 64-bit operation in which both 32-bit halves are nonzero,
+ // split the operation into two.
+ if (!ResNode && Node->getValueType(0) == MVT::i64)
+ if (ConstantSDNode *Op1 = dyn_cast<ConstantSDNode>(Node->getOperand(1))) {
+ uint64_t Val = Op1->getZExtValue();
+ if (!SystemZ::isImmLF(Val) && !SystemZ::isImmHF(Val))
+ Node = splitLargeImmediate(Opcode, Node, Node->getOperand(0),
+ Val - uint32_t(Val), uint32_t(Val));
+ }
+ break;
- #ifndef NDEBUG
- DOUT << std::string(Indent-2, ' ') << "=> ";
- if (ResNode == NULL || ResNode == Op.getNode())
- DEBUG(Op.getNode()->dump(CurDAG));
- else
- DEBUG(ResNode->dump(CurDAG));
- DOUT << "\n";
- Indent -= 2;
- #endif
+ case ISD::AND:
+ if (Node->getOperand(1).getOpcode() != ISD::Constant)
+ ResNode = tryRxSBG(Node, SystemZ::RNSBG);
+ // Fall through.
+ case ISD::ROTL:
+ case ISD::SHL:
+ case ISD::SRL:
+ if (!ResNode)
+ ResNode = tryRISBGZero(Node);
+ break;
+
+ case ISD::Constant:
+ // If this is a 64-bit constant that is out of the range of LLILF,
+ // LLIHF and LGFI, split it into two 32-bit pieces.
+ if (Node->getValueType(0) == MVT::i64) {
+ uint64_t Val = cast<ConstantSDNode>(Node)->getZExtValue();
+ if (!SystemZ::isImmLF(Val) && !SystemZ::isImmHF(Val) && !isInt<32>(Val))
+ Node = splitLargeImmediate(ISD::OR, Node, SDValue(),
+ Val - uint32_t(Val), uint32_t(Val));
+ }
+ break;
+
+ case ISD::ATOMIC_LOAD_SUB:
+ // Try to convert subtractions of constants to additions.
+ if (ConstantSDNode *Op2 = dyn_cast<ConstantSDNode>(Node->getOperand(2))) {
+ uint64_t Value = -Op2->getZExtValue();
+ EVT VT = Node->getValueType(0);
+ if (VT == MVT::i32 || isInt<32>(Value)) {
+ SDValue Ops[] = { Node->getOperand(0), Node->getOperand(1),
+ CurDAG->getConstant(int32_t(Value), VT) };
+ Node = CurDAG->MorphNodeTo(Node, ISD::ATOMIC_LOAD_ADD,
+ Node->getVTList(), Ops, array_lengthof(Ops));
+ }
+ }
+ break;
+ }
+ // Select the default instruction
+ if (!ResNode)
+ ResNode = SelectCode(Node);
+
+ DEBUG(errs() << "=> ";
+ if (ResNode == NULL || ResNode == Node)
+ Node->dump(CurDAG);
+ else
+ ResNode->dump(CurDAG);
+ errs() << "\n";
+ );
return ResNode;
}
+
+bool SystemZDAGToDAGISel::
+SelectInlineAsmMemoryOperand(const SDValue &Op,
+ char ConstraintCode,
+ std::vector<SDValue> &OutOps) {
+ assert(ConstraintCode == 'm' && "Unexpected constraint code");
+ // Accept addresses with short displacements, which are compatible
+ // with Q, R, S and T. But keep the index operand for future expansion.
+ SDValue Base, Disp, Index;
+ if (!selectBDXAddr(SystemZAddressingMode::FormBD,
+ SystemZAddressingMode::Disp12Only,
+ Op, Base, Disp, Index))
+ return true;
+ OutOps.push_back(Base);
+ OutOps.push_back(Disp);
+ OutOps.push_back(Index);
+ return false;
+}