using namespace llvm;
+#define DEBUG_TYPE "systemz-isel"
+
namespace {
// Used to build addressing modes.
struct SystemZAddressingMode {
errs() << "SystemZAddressingMode " << this << '\n';
errs() << " Base ";
- if (Base.getNode() != 0)
+ if (Base.getNode())
Base.getNode()->dump();
else
errs() << "null\n";
if (hasIndexField()) {
errs() << " Index ";
- if (Index.getNode() != 0)
+ if (Index.getNode())
Index.getNode()->dump();
else
errs() << "null\n";
return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1;
}
-// Represents operands 2 to 5 of a ROTATE AND ... SELECTED BITS operation.
-// The operands are: Input (R2), Start (I3), End (I4) and Rotate (I5).
-// The operand value is effectively (and (rotl Input Rotate) Mask) and
-// has BitSize bits.
-struct RISBGOperands {
- RISBGOperands(SDValue N)
- : BitSize(N.getValueType().getSizeInBits()), Mask(allOnes(BitSize)),
- Input(N), Start(64 - BitSize), End(63), Rotate(0) {}
-
+// 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 output value has BitSize bits, although Input may be
+// narrower (in which case the upper bits are don't care).
+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;
const SystemZSubtarget &Subtarget;
// Used by SystemZOperands.td to create integer constants.
- inline SDValue getImm(const SDNode *Node, uint64_t Imm) {
+ inline SDValue getImm(const SDNode *Node, uint64_t Imm) const {
return CurDAG->getTargetConstant(Imm, Node->getValueType(0));
}
+ const SystemZTargetMachine &getTargetMachine() const {
+ return static_cast<const SystemZTargetMachine &>(TM);
+ }
+
+ const SystemZInstrInfo *getInstrInfo() const {
+ return getTargetMachine().getInstrInfo();
+ }
+
// 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);
+ bool expandAddress(SystemZAddressingMode &AM, bool IsBase) const;
// Try to describe N in AM, returning true on success.
- bool selectAddress(SDValue N, SystemZAddressingMode &AM);
+ bool selectAddress(SDValue N, SystemZAddressingMode &AM) const;
// Extract individual target operands from matched address AM.
void getAddressOperands(const SystemZAddressingMode &AM, EVT VT,
- SDValue &Base, SDValue &Disp);
+ SDValue &Base, SDValue &Disp) const;
void getAddressOperands(const SystemZAddressingMode &AM, EVT VT,
- SDValue &Base, SDValue &Disp, SDValue &Index);
+ SDValue &Base, SDValue &Disp, SDValue &Index) const;
// 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);
+ SDValue &Base, SDValue &Disp) const;
+
+ // Try to match Addr as a FormBDX address with displacement type DR.
+ // Return true on success and if the result had no index. Store the
+ // base and displacement in Base and Disp respectively.
+ bool selectMVIAddr(SystemZAddressingMode::DispRange DR, SDValue Addr,
+ SDValue &Base, SDValue &Disp) const;
// 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);
+ SDValue &Base, SDValue &Disp, SDValue &Index) const;
// PC-relative address matching routines used by SystemZOperands.td.
- bool selectPCRelAddress(SDValue Addr, SDValue &Target) {
- if (Addr.getOpcode() == SystemZISD::PCREL_WRAPPER) {
+ bool selectPCRelAddress(SDValue Addr, SDValue &Target) const {
+ if (SystemZISD::isPCREL(Addr.getOpcode())) {
Target = Addr.getOperand(0);
return true;
}
}
// BD matching routines used by SystemZOperands.td.
- bool selectBDAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ bool selectBDAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp) const {
return selectBDAddr(SystemZAddressingMode::Disp12Only, Addr, Base, Disp);
}
- bool selectBDAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ bool selectBDAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
return selectBDAddr(SystemZAddressingMode::Disp12Pair, Addr, Base, Disp);
}
- bool selectBDAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ bool selectBDAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp) const {
return selectBDAddr(SystemZAddressingMode::Disp20Only, Addr, Base, Disp);
}
- bool selectBDAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp) {
+ bool selectBDAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
return selectBDAddr(SystemZAddressingMode::Disp20Pair, Addr, Base, Disp);
}
+ // MVI matching routines used by SystemZOperands.td.
+ bool selectMVIAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
+ return selectMVIAddr(SystemZAddressingMode::Disp12Pair, Addr, Base, Disp);
+ }
+ bool selectMVIAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp) const {
+ return selectMVIAddr(SystemZAddressingMode::Disp20Pair, Addr, Base, Disp);
+ }
+
// BDX matching routines used by SystemZOperands.td.
bool selectBDXAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp12Only,
Addr, Base, Disp, Index);
}
bool selectBDXAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp12Pair,
Addr, Base, Disp, Index);
}
bool selectDynAlloc12Only(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXDynAlloc,
SystemZAddressingMode::Disp12Only,
Addr, Base, Disp, Index);
}
bool selectBDXAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp20Only,
Addr, Base, Disp, Index);
}
bool selectBDXAddr20Only128(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp20Only128,
Addr, Base, Disp, Index);
}
bool selectBDXAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp20Pair,
Addr, Base, Disp, Index);
}
bool selectLAAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
SystemZAddressingMode::Disp12Pair,
Addr, Base, Disp, Index);
}
bool selectLAAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
- SDValue &Index) {
+ SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
SystemZAddressingMode::Disp20Pair,
Addr, Base, Disp, Index);
// 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);
+ bool detectOrAndInsertion(SDValue &Op, uint64_t InsertMask) const;
- // Try to fold some of Ops.Input into other fields of Ops. Return true
- // on success.
- bool expandRISBG(RISBGOperands &Ops);
+ // Try to update RxSBG so that only the bits of RxSBG.Input in Mask are used.
+ // Return true on success.
+ bool refineRxSBGMask(RxSBGOperands &RxSBG, uint64_t Mask) const;
- // Return an undefined i64 value.
- SDValue getUNDEF64(SDLoc DL);
+ // Try to fold some of RxSBG.Input into other fields of RxSBG.
+ // Return true on success.
+ bool expandRxSBG(RxSBGOperands &RxSBG) const;
+
+ // Return an undefined value of type VT.
+ SDValue getUNDEF(SDLoc DL, EVT VT) const;
// Convert N to VT, if it isn't already.
- SDValue convertTo(SDLoc DL, EVT VT, SDValue N);
+ SDValue convertTo(SDLoc DL, EVT VT, SDValue N) const;
// 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 ROSBG to implement OR node N. Return the selected
- // node on success, otherwise return null.
- SDNode *tryRISBGOrROSBG(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:
//
SDNode *splitLargeImmediate(unsigned Opcode, SDNode *Node, SDValue Op0,
uint64_t UpperVal, uint64_t LowerVal);
+ // Return true if Load and Store are loads and stores of the same size
+ // and are guaranteed not to overlap. Such operations can be implemented
+ // using block (SS-format) instructions.
+ //
+ // Partial overlap would lead to incorrect code, since the block operations
+ // are logically bytewise, even though they have a fast path for the
+ // non-overlapping case. We also need to avoid full overlap (i.e. two
+ // addresses that might be equal at run time) because although that case
+ // would be handled correctly, it might be implemented by millicode.
+ bool canUseBlockOperation(StoreSDNode *Store, LoadSDNode *Load) const;
+
+ // N is a (store (load Y), X) pattern. Return true if it can use an MVC
+ // from Y to X.
bool storeLoadCanUseMVC(SDNode *N) const;
+ // N is a (store (op (load A[0]), (load A[1])), X) pattern. Return true
+ // if A[1 - I] == X and if N can use a block operation like NC from A[I]
+ // to X.
+ bool storeLoadCanUseBlockBinary(SDNode *N, unsigned I) const;
+
public:
SystemZDAGToDAGISel(SystemZTargetMachine &TM, CodeGenOpt::Level OptLevel)
: SelectionDAGISel(TM, OptLevel),
Subtarget(*TM.getSubtargetImpl()) { }
// Override MachineFunctionPass.
- virtual const char *getPassName() const LLVM_OVERRIDE {
+ const char *getPassName() const override {
return "SystemZ DAG->DAG Pattern Instruction Selection";
}
// Override SelectionDAGISel.
- virtual SDNode *Select(SDNode *Node) LLVM_OVERRIDE;
- virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
- char ConstraintCode,
- std::vector<SDValue> &OutOps)
- LLVM_OVERRIDE;
+ SDNode *Select(SDNode *Node) override;
+ bool SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
+ std::vector<SDValue> &OutOps) override;
// Include the pieces autogenerated from the target description.
#include "SystemZGenDAGISel.inc"
// 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) {
+ SDValue Op0, uint64_t Op1) {
// First try adjusting the displacement.
- int64_t TestDisp = AM.Disp + Op1->getSExtValue();
+ int64_t TestDisp = AM.Disp + Op1;
if (selectDisp(AM.DR, TestDisp)) {
changeComponent(AM, IsBase, Op0);
AM.Disp = TestDisp;
}
bool SystemZDAGToDAGISel::expandAddress(SystemZAddressingMode &AM,
- bool IsBase) {
+ bool IsBase) const {
SDValue N = IsBase ? AM.Base : AM.Index;
unsigned Opcode = N.getOpcode();
if (Opcode == ISD::TRUNCATE) {
return expandAdjDynAlloc(AM, IsBase, Op0);
if (Op0Code == ISD::Constant)
- return expandDisp(AM, IsBase, Op1, cast<ConstantSDNode>(Op0));
+ return expandDisp(AM, IsBase, Op1,
+ cast<ConstantSDNode>(Op0)->getSExtValue());
if (Op1Code == ISD::Constant)
- return expandDisp(AM, IsBase, Op0, cast<ConstantSDNode>(Op1));
+ return expandDisp(AM, IsBase, Op0,
+ cast<ConstantSDNode>(Op1)->getSExtValue());
if (IsBase && expandIndex(AM, Op0, Op1))
return true;
}
+ if (Opcode == SystemZISD::PCREL_OFFSET) {
+ SDValue Full = N.getOperand(0);
+ SDValue Base = N.getOperand(1);
+ SDValue Anchor = Base.getOperand(0);
+ uint64_t Offset = (cast<GlobalAddressSDNode>(Full)->getOffset() -
+ cast<GlobalAddressSDNode>(Anchor)->getOffset());
+ return expandDisp(AM, IsBase, Base, Offset);
+ }
return false;
}
// Return true if Addr is suitable for AM, updating AM if so.
bool SystemZDAGToDAGISel::selectAddress(SDValue Addr,
- SystemZAddressingMode &AM) {
+ SystemZAddressingMode &AM) const {
// 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)))
+ expandDisp(AM, true, SDValue(),
+ cast<ConstantSDNode>(Addr)->getSExtValue()))
;
else
// Otherwise try expanding each component.
void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
EVT VT, SDValue &Base,
- SDValue &Disp) {
+ SDValue &Disp) const {
Base = AM.Base;
if (!Base.getNode())
// Register 0 means "no base". This is mostly useful for shifts.
void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
EVT VT, SDValue &Base,
- SDValue &Disp, SDValue &Index) {
+ SDValue &Disp,
+ SDValue &Index) const {
getAddressOperands(AM, VT, Base, Disp);
Index = AM.Index;
bool SystemZDAGToDAGISel::selectBDAddr(SystemZAddressingMode::DispRange DR,
SDValue Addr, SDValue &Base,
- SDValue &Disp) {
+ SDValue &Disp) const {
SystemZAddressingMode AM(SystemZAddressingMode::FormBD, DR);
if (!selectAddress(Addr, AM))
return false;
return true;
}
+bool SystemZDAGToDAGISel::selectMVIAddr(SystemZAddressingMode::DispRange DR,
+ SDValue Addr, SDValue &Base,
+ SDValue &Disp) const {
+ SystemZAddressingMode AM(SystemZAddressingMode::FormBDXNormal, DR);
+ if (!selectAddress(Addr, AM) || AM.Index.getNode())
+ 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) {
+ SDValue &Disp, SDValue &Index) const {
SystemZAddressingMode AM(Form, DR);
if (!selectAddress(Addr, AM))
return false;
}
bool SystemZDAGToDAGISel::detectOrAndInsertion(SDValue &Op,
- uint64_t InsertMask) {
+ uint64_t InsertMask) const {
// 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());
+ auto *MaskNode = dyn_cast<ConstantSDNode>(Op.getOperand(1).getNode());
if (!MaskNode)
return false;
uint64_t Used = allOnes(Op.getValueType().getSizeInBits());
if (Used != (AndMask | InsertMask)) {
APInt KnownZero, KnownOne;
- CurDAG->ComputeMaskedBits(Op.getOperand(0), KnownZero, KnownOne);
+ CurDAG->computeKnownBits(Op.getOperand(0), KnownZero, KnownOne);
if (Used != (AndMask | InsertMask | KnownZero.getZExtValue()))
return false;
}
return true;
}
-// 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;
- }
+bool SystemZDAGToDAGISel::refineRxSBGMask(RxSBGOperands &RxSBG,
+ uint64_t Mask) const {
+ const SystemZInstrInfo *TII = getInstrInfo();
+ if (RxSBG.Rotate != 0)
+ Mask = (Mask << RxSBG.Rotate) | (Mask >> (64 - RxSBG.Rotate));
+ Mask &= RxSBG.Mask;
+ if (TII->isRxSBGMask(Mask, RxSBG.BitSize, RxSBG.Start, RxSBG.End)) {
+ RxSBG.Mask = Mask;
+ return true;
+ }
return false;
}
-// Try to update RISBG so that only the bits of Ops.Input in Mask are used.
-// Return true on success.
-static bool refineRISBGMask(RISBGOperands &RISBG, uint64_t Mask) {
- if (RISBG.Rotate != 0)
- Mask = (Mask << RISBG.Rotate) | (Mask >> (64 - RISBG.Rotate));
- Mask &= RISBG.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))
- {
- RISBG.Mask = Mask;
- RISBG.Start = 63 - (LSB + Length - 1);
- RISBG.End = 63 - LSB;
- return true;
- }
-
- // 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(RISBG.BitSize), LSB, Length))
- {
- assert(LSB > 0 && "Bottom bit must be set");
- assert(LSB + Length < RISBG.BitSize && "Top bit must be set");
- RISBG.Mask = Mask;
- RISBG.Start = 63 - (LSB - 1);
- RISBG.End = 63 - (LSB + Length);
- return true;
- }
-
- return false;
+// Return true if any bits of (RxSBG.Input & Mask) are significant.
+static bool maskMatters(RxSBGOperands &RxSBG, uint64_t Mask) {
+ // Rotate the mask in the same way as RxSBG.Input is rotated.
+ if (RxSBG.Rotate != 0)
+ Mask = ((Mask << RxSBG.Rotate) | (Mask >> (64 - RxSBG.Rotate)));
+ return (Mask & RxSBG.Mask) != 0;
}
-bool SystemZDAGToDAGISel::expandRISBG(RISBGOperands &RISBG) {
- SDValue N = RISBG.Input;
- switch (N.getOpcode()) {
+bool SystemZDAGToDAGISel::expandRxSBG(RxSBGOperands &RxSBG) const {
+ SDValue N = RxSBG.Input;
+ unsigned Opcode = N.getOpcode();
+ switch (Opcode) {
case ISD::AND: {
- ConstantSDNode *MaskNode =
- dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ if (RxSBG.Opcode == SystemZ::RNSBG)
+ return false;
+
+ auto *MaskNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
if (!MaskNode)
return false;
SDValue Input = N.getOperand(0);
uint64_t Mask = MaskNode->getZExtValue();
- if (!refineRISBGMask(RISBG, Mask)) {
+ 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);
+ CurDAG->computeKnownBits(Input, KnownZero, KnownOne);
Mask |= KnownZero.getZExtValue();
- if (!refineRISBGMask(RISBG, Mask))
+ if (!refineRxSBGMask(RxSBG, Mask))
return false;
}
- RISBG.Input = Input;
+ RxSBG.Input = Input;
+ return true;
+ }
+
+ case ISD::OR: {
+ if (RxSBG.Opcode != SystemZ::RNSBG)
+ return false;
+
+ auto *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->computeKnownBits(Input, KnownZero, KnownOne);
+ Mask &= ~KnownOne.getZExtValue();
+ if (!refineRxSBGMask(RxSBG, Mask))
+ return false;
+ }
+ RxSBG.Input = Input;
return true;
}
case ISD::ROTL: {
- // Any 64-bit rotate left can be merged into the RISBG.
- if (RISBG.BitSize != 64)
+ // Any 64-bit rotate left can be merged into the RxSBG.
+ if (RxSBG.BitSize != 64 || N.getValueType() != MVT::i64)
return false;
- ConstantSDNode *CountNode
- = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
if (!CountNode)
return false;
- RISBG.Rotate = (RISBG.Rotate + CountNode->getZExtValue()) & 63;
- RISBG.Input = N.getOperand(0);
+ RxSBG.Rotate = (RxSBG.Rotate + CountNode->getZExtValue()) & 63;
+ RxSBG.Input = N.getOperand(0);
return true;
}
- case ISD::SHL: {
- // Treat (shl X, count) as (and (rotl X, count), ~0<<count).
- ConstantSDNode *CountNode =
- dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
- if (!CountNode)
- return false;
+ case ISD::ANY_EXTEND:
+ // Bits above the extended operand are don't-care.
+ RxSBG.Input = N.getOperand(0);
+ return true;
- uint64_t Count = CountNode->getZExtValue();
- if (Count < 1 ||
- Count >= RISBG.BitSize ||
- !refineRISBGMask(RISBG, allOnes(RISBG.BitSize - Count) << Count))
+ case ISD::ZERO_EXTEND:
+ if (RxSBG.Opcode != SystemZ::RNSBG) {
+ // Restrict the mask to the extended operand.
+ unsigned InnerBitSize = N.getOperand(0).getValueType().getSizeInBits();
+ if (!refineRxSBGMask(RxSBG, allOnes(InnerBitSize)))
+ return false;
+
+ RxSBG.Input = N.getOperand(0);
+ return true;
+ }
+ // Fall through.
+
+ case ISD::SIGN_EXTEND: {
+ // Check that the extension bits are don't-care (i.e. are masked out
+ // by the final mask).
+ unsigned InnerBitSize = N.getOperand(0).getValueType().getSizeInBits();
+ if (maskMatters(RxSBG, allOnes(RxSBG.BitSize) - allOnes(InnerBitSize)))
return false;
- RISBG.Rotate = (RISBG.Rotate + Count) & 63;
- RISBG.Input = N.getOperand(0);
+ RxSBG.Input = N.getOperand(0);
return true;
}
- case ISD::SRL: {
- // Treat (srl X, count), mask) as (and (rotl X, size-count), ~0>>count),
- // which is similar to SLL above.
- ConstantSDNode *CountNode =
- dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ case ISD::SHL: {
+ auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
if (!CountNode)
return false;
uint64_t Count = CountNode->getZExtValue();
- if (Count < 1 ||
- Count >= RISBG.BitSize ||
- !refineRISBGMask(RISBG, allOnes(RISBG.BitSize - Count)))
+ unsigned BitSize = N.getValueType().getSizeInBits();
+ if (Count < 1 || Count >= BitSize)
return false;
- RISBG.Rotate = (RISBG.Rotate - Count) & 63;
- RISBG.Input = N.getOperand(0);
+ 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 (maskMatters(RxSBG, allOnes(Count)))
+ return false;
+ } else {
+ // Treat (shl X, count) as (and (rotl X, count), ~0<<count).
+ if (!refineRxSBGMask(RxSBG, allOnes(BitSize - Count) << Count))
+ return false;
+ }
+
+ RxSBG.Rotate = (RxSBG.Rotate + Count) & 63;
+ RxSBG.Input = N.getOperand(0);
return true;
}
+ case ISD::SRL:
case ISD::SRA: {
- // Treat (sra X, count) as (rotl X, size-count) as long as the top
- // count bits from Ops.Input are ignored.
- ConstantSDNode *CountNode =
- dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
+ auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
if (!CountNode)
return false;
uint64_t Count = CountNode->getZExtValue();
- if (RISBG.Rotate != 0 ||
- Count < 1 ||
- Count >= RISBG.BitSize ||
- RISBG.Start < 64 - (RISBG.BitSize - Count))
+ unsigned BitSize = N.getValueType().getSizeInBits();
+ if (Count < 1 || Count >= BitSize)
return false;
- RISBG.Rotate = -Count & 63;
- RISBG.Input = N.getOperand(0);
+ 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 (maskMatters(RxSBG, allOnes(Count) << (BitSize - Count)))
+ 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(BitSize - Count)))
+ return false;
+ }
+
+ RxSBG.Rotate = (RxSBG.Rotate - Count) & 63;
+ RxSBG.Input = N.getOperand(0);
return true;
}
default:
}
}
-SDValue SystemZDAGToDAGISel::getUNDEF64(SDLoc DL) {
- SDNode *N = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, MVT::i64);
+SDValue SystemZDAGToDAGISel::getUNDEF(SDLoc DL, EVT VT) const {
+ SDNode *N = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, VT);
return SDValue(N, 0);
}
-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);
- }
+SDValue SystemZDAGToDAGISel::convertTo(SDLoc DL, EVT VT, SDValue N) const {
+ if (N.getValueType() == MVT::i32 && VT == MVT::i64)
+ return CurDAG->getTargetInsertSubreg(SystemZ::subreg_l32,
+ DL, VT, getUNDEF(DL, MVT::i64), N);
+ if (N.getValueType() == MVT::i64 && VT == MVT::i32)
+ return CurDAG->getTargetExtractSubreg(SystemZ::subreg_l32, DL, VT, N);
assert(N.getValueType() == VT && "Unexpected value types");
return N;
}
SDNode *SystemZDAGToDAGISel::tryRISBGZero(SDNode *N) {
- RISBGOperands RISBG(SDValue(N, 0));
- unsigned Count = 0;
- while (expandRISBG(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);
+ RxSBGOperands RISBG(SystemZ::RISBG, SDValue(N, 0));
+ unsigned Count = 0;
+ while (expandRxSBG(RISBG))
+ if (RISBG.Input.getOpcode() != ISD::ANY_EXTEND)
+ Count += 1;
+ if (Count == 0)
+ return nullptr;
+ if (Count == 1) {
+ // Prefer to use normal shift instructions over RISBG, since they can handle
+ // all cases and are sometimes shorter.
+ if (N->getOpcode() != ISD::AND)
+ return nullptr;
+
+ // Prefer register extensions like LLC over RISBG. Also prefer to start
+ // out with normal ANDs if one instruction would be enough. We can convert
+ // these ANDs into an RISBG later if a three-address instruction is useful.
+ if (VT == MVT::i32 ||
+ RISBG.Mask == 0xff ||
+ RISBG.Mask == 0xffff ||
+ SystemZ::isImmLF(~RISBG.Mask) ||
+ SystemZ::isImmHF(~RISBG.Mask)) {
+ // Force the new mask into the DAG, since it may include known-one bits.
+ auto *MaskN = cast<ConstantSDNode>(N->getOperand(1).getNode());
+ if (MaskN->getZExtValue() != RISBG.Mask) {
+ SDValue NewMask = CurDAG->getConstant(RISBG.Mask, VT);
+ N = CurDAG->UpdateNodeOperands(N, N->getOperand(0), NewMask);
+ return SelectCode(N);
+ }
+ return nullptr;
+ }
+ }
+
+ unsigned Opcode = SystemZ::RISBG;
+ EVT OpcodeVT = MVT::i64;
+ if (VT == MVT::i32 && Subtarget.hasHighWord()) {
+ Opcode = SystemZ::RISBMux;
+ OpcodeVT = MVT::i32;
+ RISBG.Start &= 31;
+ RISBG.End &= 31;
+ }
SDValue Ops[5] = {
- getUNDEF64(SDLoc(N)),
- convertTo(SDLoc(N), MVT::i64, RISBG.Input),
+ getUNDEF(SDLoc(N), OpcodeVT),
+ convertTo(SDLoc(N), OpcodeVT, 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);
+ N = CurDAG->getMachineNode(Opcode, SDLoc(N), OpcodeVT, Ops);
return convertTo(SDLoc(N), VT, SDValue(N, 0)).getNode();
}
-SDNode *SystemZDAGToDAGISel::tryRISBGOrROSBG(SDNode *N) {
- // Try treating each operand of N as the second operand of RISBG or ROSBG
+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.
- RISBGOperands RISBG[] = { N->getOperand(0), N->getOperand(1) };
+ RxSBGOperands RxSBG[] = {
+ RxSBGOperands(Opcode, N->getOperand(0)),
+ RxSBGOperands(Opcode, N->getOperand(1))
+ };
unsigned Count[] = { 0, 0 };
for (unsigned I = 0; I < 2; ++I)
- while (expandRISBG(RISBG[I]))
- Count[I] += 1;
+ while (expandRxSBG(RxSBG[I]))
+ if (RxSBG[I].Input.getOpcode() != ISD::ANY_EXTEND)
+ Count[I] += 1;
// Do nothing if neither operand is suitable.
if (Count[0] == 0 && Count[1] == 0)
- return 0;
+ return nullptr;
// 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 ((RISBG[I].Mask & 0xff) == 0)
- if (LoadSDNode *Load = dyn_cast<LoadSDNode>(Op0.getNode()))
+ if (Opcode == SystemZ::ROSBG && (RxSBG[I].Mask & 0xff) == 0)
+ if (auto *Load = dyn_cast<LoadSDNode>(Op0.getNode()))
if (Load->getMemoryVT() == MVT::i8)
- return 0;
+ return nullptr;
// See whether we can avoid an AND in the first operand by converting
// ROSBG to RISBG.
- unsigned Opcode = SystemZ::ROSBG;
- if (detectOrAndInsertion(Op0, RISBG[I].Mask))
+ 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, RISBG[I].Input),
- CurDAG->getTargetConstant(RISBG[I].Start, MVT::i32),
- CurDAG->getTargetConstant(RISBG[I].End, MVT::i32),
- CurDAG->getTargetConstant(RISBG[I].Rotate, MVT::i32)
+ 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 Or.getNode();
}
-// 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());
+bool SystemZDAGToDAGISel::canUseBlockOperation(StoreSDNode *Store,
+ LoadSDNode *Load) const {
+ // Check that the two memory operands have the same size.
+ if (Load->getMemoryVT() != Store->getMemoryVT())
+ return false;
- // MVC is logically a bytewise copy, so can't be used for volatile accesses.
+ // Volatility stops an access from being decomposed.
if (Load->isVolatile() || Store->isVolatile())
return false;
+ // There's no chance of overlap if the load is invariant.
+ if (Load->isInvariant())
+ return true;
+
+ // Otherwise we need to check whether there's an alias.
+ const Value *V1 = Load->getMemOperand()->getValue();
+ const Value *V2 = Store->getMemOperand()->getValue();
+ if (!V1 || !V2)
+ return false;
+
+ // Reject equality.
+ uint64_t Size = Load->getMemoryVT().getStoreSize();
+ int64_t End1 = Load->getSrcValueOffset() + Size;
+ int64_t End2 = Store->getSrcValueOffset() + Size;
+ if (V1 == V2 && End1 == End2)
+ return false;
+
+ return !AA->alias(AliasAnalysis::Location(V1, End1, Load->getAAInfo()),
+ AliasAnalysis::Location(V2, End2, Store->getAAInfo()));
+}
+
+bool SystemZDAGToDAGISel::storeLoadCanUseMVC(SDNode *N) const {
+ auto *Store = cast<StoreSDNode>(N);
+ auto *Load = cast<LoadSDNode>(Store->getValue());
+
// 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)
+ if (SystemZISD::isPCREL(Load->getBasePtr().getOpcode()))
return false;
// Prefer STHRL, STRL and STGRL.
- if (Store->getBasePtr().getOpcode() == SystemZISD::PCREL_WRAPPER)
+ if (SystemZISD::isPCREL(Store->getBasePtr().getOpcode()))
return false;
}
- // There's no chance of overlap if the load is invariant.
- if (Load->isInvariant())
- return true;
-
- // If both operands are aligned, they must be equal or not overlap.
- if (Load->getAlignment() >= Size && Store->getAlignment() >= Size)
- return true;
-
- // 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;
+ return canUseBlockOperation(Store, Load);
+}
- 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()));
+bool SystemZDAGToDAGISel::storeLoadCanUseBlockBinary(SDNode *N,
+ unsigned I) const {
+ auto *StoreA = cast<StoreSDNode>(N);
+ auto *LoadA = cast<LoadSDNode>(StoreA->getValue().getOperand(1 - I));
+ auto *LoadB = cast<LoadSDNode>(StoreA->getValue().getOperand(I));
+ return !LoadA->isVolatile() && canUseBlockOperation(StoreA, LoadB);
}
SDNode *SystemZDAGToDAGISel::Select(SDNode *Node) {
// If we have a custom node, we already have selected!
if (Node->isMachineOpcode()) {
DEBUG(errs() << "== "; Node->dump(CurDAG); errs() << "\n");
- return 0;
+ Node->setNodeId(-1);
+ return nullptr;
}
unsigned Opcode = Node->getOpcode();
- SDNode *ResNode = 0;
+ SDNode *ResNode = nullptr;
switch (Opcode) {
case ISD::OR:
if (Node->getOperand(1).getOpcode() != ISD::Constant)
- ResNode = tryRISBGOrROSBG(Node);
- // Fall through.
+ 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))) {
+ if (auto *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),
break;
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:
- ResNode = tryRISBGZero(Node);
+ case ISD::ZERO_EXTEND:
+ if (!ResNode)
+ ResNode = tryRISBGZero(Node);
break;
case ISD::Constant:
}
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));
- }
+ case SystemZISD::SELECT_CCMASK: {
+ SDValue Op0 = Node->getOperand(0);
+ SDValue Op1 = Node->getOperand(1);
+ // Prefer to put any load first, so that it can be matched as a
+ // conditional load.
+ if (Op1.getOpcode() == ISD::LOAD && Op0.getOpcode() != ISD::LOAD) {
+ SDValue CCValid = Node->getOperand(2);
+ SDValue CCMask = Node->getOperand(3);
+ uint64_t ConstCCValid =
+ cast<ConstantSDNode>(CCValid.getNode())->getZExtValue();
+ uint64_t ConstCCMask =
+ cast<ConstantSDNode>(CCMask.getNode())->getZExtValue();
+ // Invert the condition.
+ CCMask = CurDAG->getConstant(ConstCCValid ^ ConstCCMask,
+ CCMask.getValueType());
+ SDValue Op4 = Node->getOperand(4);
+ Node = CurDAG->UpdateNodeOperands(Node, Op1, Op0, CCValid, CCMask, Op4);
}
break;
}
+ }
// Select the default instruction
if (!ResNode)
ResNode = SelectCode(Node);
DEBUG(errs() << "=> ";
- if (ResNode == NULL || ResNode == Node)
+ if (ResNode == nullptr || ResNode == Node)
Node->dump(CurDAG);
else
ResNode->dump(CurDAG);
projects / oota-llvm.git / blobdiff