maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/lib/Target/SystemZ/SystemZISelDAGToDAG.cpp

1847 lines
66 KiB
C++
Raw Blame History

//===-- SystemZISelDAGToDAG.cpp - A dag to dag inst selector for SystemZ --===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines an instruction selector for the SystemZ target.
//
//===----------------------------------------------------------------------===//
#include "SystemZTargetMachine.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
#define DEBUG_TYPE "systemz-isel"
namespace {
// Used to build addressing modes.
struct SystemZAddressingMode {
// The shape of the address.
enum AddrForm {
// base+displacement
FormBD,
// 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;
// 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;
SystemZAddressingMode(AddrForm form, DispRange dr)
: Form(form), DR(dr), Base(), Disp(0), Index(),
IncludesDynAlloc(false) {}
// True if the address can have an index register.
bool hasIndexField() { return Form != FormBD; }
// True if the address can (and must) include ADJDYNALLOC.
bool isDynAlloc() { return Form == FormBDXDynAlloc; }
void dump(const llvm::SelectionDAG *DAG) {
errs() << "SystemZAddressingMode " << this << '\n';
errs() << " Base ";
if (Base.getNode())
Base.getNode()->dump(DAG);
else
errs() << "null\n";
if (hasIndexField()) {
errs() << " Index ";
if (Index.getNode())
Index.getNode()->dump(DAG);
else
errs() << "null\n";
}
errs() << " Disp " << Disp;
if (IncludesDynAlloc)
errs() << " + ADJDYNALLOC";
errs() << '\n';
}
};
// Return a mask with Count low bits set.
static uint64_t allOnes(unsigned int Count) {
assert(Count <= 64);
if (Count > 63)
return UINT64_MAX;
return (uint64_t(1) << Count) - 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 output value has BitSize bits, although Input may be
// narrower (in which case the upper bits are don't care), or wider (in which
// case the result will be truncated as part of the operation).
struct RxSBGOperands {
RxSBGOperands(unsigned Op, SDValue N)
: Opcode(Op), BitSize(N.getValueSizeInBits()),
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 SystemZSubtarget *Subtarget;
// Used by SystemZOperands.td to create integer constants.
inline SDValue getImm(const SDNode *Node, uint64_t Imm) const {
return CurDAG->getTargetConstant(Imm, SDLoc(Node), Node->getValueType(0));
}
const SystemZTargetMachine &getTargetMachine() const {
return static_cast<const SystemZTargetMachine &>(TM);
}
const SystemZInstrInfo *getInstrInfo() const {
return Subtarget->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) const;
// Try to describe N in AM, returning true on success.
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) const;
void getAddressOperands(const SystemZAddressingMode &AM, EVT VT,
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) 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) const;
// PC-relative address matching routines used by SystemZOperands.td.
bool selectPCRelAddress(SDValue Addr, SDValue &Target) const {
if (SystemZISD::isPCREL(Addr.getOpcode())) {
Target = Addr.getOperand(0);
return true;
}
return false;
}
// BD matching routines used by SystemZOperands.td.
bool selectBDAddr12Only(SDValue Addr, SDValue &Base, SDValue &Disp) const {
return selectBDAddr(SystemZAddressingMode::Disp12Only, Addr, Base, 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) const {
return selectBDAddr(SystemZAddressingMode::Disp20Only, Addr, Base, 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) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp12Only,
Addr, Base, Disp, Index);
}
bool selectBDXAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp12Pair,
Addr, Base, Disp, Index);
}
bool selectDynAlloc12Only(SDValue Addr, SDValue &Base, SDValue &Disp,
SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXDynAlloc,
SystemZAddressingMode::Disp12Only,
Addr, Base, Disp, Index);
}
bool selectBDXAddr20Only(SDValue Addr, SDValue &Base, SDValue &Disp,
SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp20Only,
Addr, Base, Disp, Index);
}
bool selectBDXAddr20Only128(SDValue Addr, SDValue &Base, SDValue &Disp,
SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp20Only128,
Addr, Base, Disp, Index);
}
bool selectBDXAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXNormal,
SystemZAddressingMode::Disp20Pair,
Addr, Base, Disp, Index);
}
bool selectLAAddr12Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
SystemZAddressingMode::Disp12Pair,
Addr, Base, Disp, Index);
}
bool selectLAAddr20Pair(SDValue Addr, SDValue &Base, SDValue &Disp,
SDValue &Index) const {
return selectBDXAddr(SystemZAddressingMode::FormBDXLA,
SystemZAddressingMode::Disp20Pair,
Addr, Base, Disp, Index);
}
// Try to match Addr as an address with a base, 12-bit displacement
// and index, where the index is element Elem of a vector.
// Return true on success, storing the base, displacement and vector
// in Base, Disp and Index respectively.
bool selectBDVAddr12Only(SDValue Addr, SDValue Elem, SDValue &Base,
SDValue &Disp, SDValue &Index) const;
// 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) const;
// 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;
// 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(const SDLoc &DL, EVT VT) const;
// Convert N to VT, if it isn't already.
SDValue convertTo(const 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.
bool 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.
bool 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)
void splitLargeImmediate(unsigned Opcode, SDNode *Node, SDValue Op0,
uint64_t UpperVal, uint64_t LowerVal);
// Try to use gather instruction Opcode to implement vector insertion N.
bool tryGather(SDNode *N, unsigned Opcode);
// Try to use scatter instruction Opcode to implement store Store.
bool tryScatter(StoreSDNode *Store, unsigned Opcode);
// Change a chain of {load; op; store} of the same value into a simple op
// through memory of that value, if the uses of the modified value and its
// address are suitable.
bool tryFoldLoadStoreIntoMemOperand(SDNode *Node);
// 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;
// Try to expand a boolean SELECT_CCMASK using an IPM sequence.
SDValue expandSelectBoolean(SDNode *Node);
public:
SystemZDAGToDAGISel(SystemZTargetMachine &TM, CodeGenOpt::Level OptLevel)
: SelectionDAGISel(TM, OptLevel) {}
bool runOnMachineFunction(MachineFunction &MF) override {
Subtarget = &MF.getSubtarget<SystemZSubtarget>();
return SelectionDAGISel::runOnMachineFunction(MF);
}
// Override MachineFunctionPass.
StringRef getPassName() const override {
return "SystemZ DAG->DAG Pattern Instruction Selection";
}
// Override SelectionDAGISel.
void Select(SDNode *Node) override;
bool SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
std::vector<SDValue> &OutOps) override;
bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
void PreprocessISelDAG() override;
// Include the pieces autogenerated from the target description.
#include "SystemZGenDAGISel.inc"
};
} // end anonymous namespace
FunctionPass *llvm::createSystemZISelDag(SystemZTargetMachine &TM,
CodeGenOpt::Level OptLevel) {
return new SystemZDAGToDAGISel(TM, OptLevel);
}
// 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;
}
// 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;
}
// 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, uint64_t Op1) {
// First try adjusting the displacement.
int64_t TestDisp = AM.Disp + Op1;
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;
}
bool SystemZDAGToDAGISel::expandAddress(SystemZAddressingMode &AM,
bool IsBase) const {
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)->getSExtValue());
if (Op1Code == ISD::Constant)
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 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;
}
// 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) 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)->getSExtValue()))
;
// Also see if it's a bare ADJDYNALLOC.
else if (Addr.getOpcode() == SystemZISD::ADJDYNALLOC &&
expandAdjDynAlloc(AM, true, SDValue()))
;
else
// 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;
LLVM_DEBUG(AM.dump(CurDAG));
return true;
}
// 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->getNodeId() == -1 ||
(SelectionDAGISel::getUninvalidatedNodeId(N.getNode()) >
SelectionDAGISel::getUninvalidatedNodeId(Pos))) {
DAG->RepositionNode(Pos->getIterator(), N.getNode());
// Mark Node as invalid for pruning as after this it may be a successor to a
// selected node but otherwise be in the same position of Pos.
// Conservatively mark it with the same -abs(Id) to assure node id
// invariant is preserved.
N->setNodeId(Pos->getNodeId());
SelectionDAGISel::InvalidateNodeId(N.getNode());
}
}
void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
EVT VT, SDValue &Base,
SDValue &Disp) const {
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;
}
// Lower the displacement to a TargetConstant.
Disp = CurDAG->getTargetConstant(AM.Disp, SDLoc(Base), VT);
}
void SystemZDAGToDAGISel::getAddressOperands(const SystemZAddressingMode &AM,
EVT VT, SDValue &Base,
SDValue &Disp,
SDValue &Index) const {
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) const {
SystemZAddressingMode AM(SystemZAddressingMode::FormBD, DR);
if (!selectAddress(Addr, AM))
return false;
getAddressOperands(AM, Addr.getValueType(), Base, Disp);
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) const {
SystemZAddressingMode AM(Form, DR);
if (!selectAddress(Addr, AM))
return false;
getAddressOperands(AM, Addr.getValueType(), Base, Disp, Index);
return true;
}
bool SystemZDAGToDAGISel::selectBDVAddr12Only(SDValue Addr, SDValue Elem,
SDValue &Base,
SDValue &Disp,
SDValue &Index) const {
SDValue Regs[2];
if (selectBDXAddr12Only(Addr, Regs[0], Disp, Regs[1]) &&
Regs[0].getNode() && Regs[1].getNode()) {
for (unsigned int I = 0; I < 2; ++I) {
Base = Regs[I];
Index = Regs[1 - I];
// We can't tell here whether the index vector has the right type
// for the access; the caller needs to do that instead.
if (Index.getOpcode() == ISD::ZERO_EXTEND)
Index = Index.getOperand(0);
if (Index.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
Index.getOperand(1) == Elem) {
Index = Index.getOperand(0);
return true;
}
}
}
return false;
}
bool SystemZDAGToDAGISel::detectOrAndInsertion(SDValue &Op,
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.
auto *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.getValueSizeInBits());
if (Used != (AndMask | InsertMask)) {
KnownBits Known;
CurDAG->computeKnownBits(Op.getOperand(0), Known);
if (Used != (AndMask | InsertMask | Known.Zero.getZExtValue()))
return false;
}
Op = Op.getOperand(0);
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;
}
// 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::expandRxSBG(RxSBGOperands &RxSBG) const {
SDValue N = RxSBG.Input;
unsigned Opcode = N.getOpcode();
switch (Opcode) {
case ISD::TRUNCATE: {
if (RxSBG.Opcode == SystemZ::RNSBG)
return false;
uint64_t BitSize = N.getValueSizeInBits();
uint64_t Mask = allOnes(BitSize);
if (!refineRxSBGMask(RxSBG, Mask))
return false;
RxSBG.Input = N.getOperand(0);
return true;
}
case ISD::AND: {
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 zeros, those bits will have
// been removed from the mask. See if adding them back in makes the
// mask suitable.
KnownBits Known;
CurDAG->computeKnownBits(Input, Known);
Mask |= Known.Zero.getZExtValue();
if (!refineRxSBGMask(RxSBG, Mask))
return false;
}
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.
KnownBits Known;
CurDAG->computeKnownBits(Input, Known);
Mask &= ~Known.One.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 RxSBG.
if (RxSBG.BitSize != 64 || N.getValueType() != MVT::i64)
return false;
auto *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::ANY_EXTEND:
// Bits above the extended operand are don't-care.
RxSBG.Input = N.getOperand(0);
return true;
case ISD::ZERO_EXTEND:
if (RxSBG.Opcode != SystemZ::RNSBG) {
// Restrict the mask to the extended operand.
unsigned InnerBitSize = N.getOperand(0).getValueSizeInBits();
if (!refineRxSBGMask(RxSBG, allOnes(InnerBitSize)))
return false;
RxSBG.Input = N.getOperand(0);
return true;
}
LLVM_FALLTHROUGH;
case ISD::SIGN_EXTEND: {
// Check that the extension bits are don't-care (i.e. are masked out
// by the final mask).
unsigned BitSize = N.getValueSizeInBits();
unsigned InnerBitSize = N.getOperand(0).getValueSizeInBits();
if (maskMatters(RxSBG, allOnes(BitSize) - allOnes(InnerBitSize))) {
// In the case where only the sign bit is active, increase Rotate with
// the extension width.
if (RxSBG.Mask == 1 && RxSBG.Rotate == 1)
RxSBG.Rotate += (BitSize - InnerBitSize);
else
return false;
}
RxSBG.Input = N.getOperand(0);
return true;
}
case ISD::SHL: {
auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
if (!CountNode)
return false;
uint64_t Count = CountNode->getZExtValue();
unsigned BitSize = N.getValueSizeInBits();
if (Count < 1 || Count >= 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 (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: {
auto *CountNode = dyn_cast<ConstantSDNode>(N.getOperand(1).getNode());
if (!CountNode)
return false;
uint64_t Count = CountNode->getZExtValue();
unsigned BitSize = N.getValueSizeInBits();
if (Count < 1 || Count >= 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 (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:
return false;
}
}
SDValue SystemZDAGToDAGISel::getUNDEF(const SDLoc &DL, EVT VT) const {
SDNode *N = CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF, DL, VT);
return SDValue(N, 0);
}
SDValue SystemZDAGToDAGISel::convertTo(const 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;
}
bool SystemZDAGToDAGISel::tryRISBGZero(SDNode *N) {
SDLoc DL(N);
EVT VT = N->getValueType(0);
if (!VT.isInteger() || VT.getSizeInBits() > 64)
return false;
RxSBGOperands RISBG(SystemZ::RISBG, SDValue(N, 0));
unsigned Count = 0;
while (expandRxSBG(RISBG))
// The widening or narrowing is expected to be free.
// Counting widening or narrowing as a saved operation will result in
// preferring an R*SBG over a simple shift/logical instruction.
if (RISBG.Input.getOpcode() != ISD::ANY_EXTEND &&
RISBG.Input.getOpcode() != ISD::TRUNCATE)
Count += 1;
if (Count == 0)
return false;
// Prefer to use normal shift instructions over RISBG, since they can handle
// all cases and are sometimes shorter.
if (Count == 1 && N->getOpcode() != ISD::AND)
return false;
// 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 (RISBG.Rotate == 0) {
bool PreferAnd = false;
// Prefer AND for any 32-bit and-immediate operation.
if (VT == MVT::i32)
PreferAnd = true;
// As well as for any 64-bit operation that can be implemented via LLC(R),
// LLH(R), LLGT(R), or one of the and-immediate instructions.
else if (RISBG.Mask == 0xff ||
RISBG.Mask == 0xffff ||
RISBG.Mask == 0x7fffffff ||
SystemZ::isImmLF(~RISBG.Mask) ||
SystemZ::isImmHF(~RISBG.Mask))
PreferAnd = true;
// And likewise for the LLZRGF instruction, which doesn't have a register
// to register version.
else if (auto *Load = dyn_cast<LoadSDNode>(RISBG.Input)) {
if (Load->getMemoryVT() == MVT::i32 &&
(Load->getExtensionType() == ISD::EXTLOAD ||
Load->getExtensionType() == ISD::ZEXTLOAD) &&
RISBG.Mask == 0xffffff00 &&
Subtarget->hasLoadAndZeroRightmostByte())
PreferAnd = true;
}
if (PreferAnd) {
// Replace the current node with an AND. Note that the current node
// might already be that same AND, in which case it is already CSE'd
// with it, and we must not call ReplaceNode.
SDValue In = convertTo(DL, VT, RISBG.Input);
SDValue Mask = CurDAG->getConstant(RISBG.Mask, DL, VT);
SDValue New = CurDAG->getNode(ISD::AND, DL, VT, In, Mask);
if (N != New.getNode()) {
insertDAGNode(CurDAG, N, Mask);
insertDAGNode(CurDAG, N, New);
ReplaceNode(N, New.getNode());
N = New.getNode();
}
// Now, select the machine opcode to implement this operation.
if (!N->isMachineOpcode())
SelectCode(N);
return true;
}
}
unsigned Opcode = SystemZ::RISBG;
// Prefer RISBGN if available, since it does not clobber CC.
if (Subtarget->hasMiscellaneousExtensions())
Opcode = SystemZ::RISBGN;
EVT OpcodeVT = MVT::i64;
if (VT == MVT::i32 && Subtarget->hasHighWord() &&
// We can only use the 32-bit instructions if all source bits are
// in the low 32 bits without wrapping, both after rotation (because
// of the smaller range for Start and End) and before rotation
// (because the input value is truncated).
RISBG.Start >= 32 && RISBG.End >= RISBG.Start &&
((RISBG.Start + RISBG.Rotate) & 63) >= 32 &&
((RISBG.End + RISBG.Rotate) & 63) >=
((RISBG.Start + RISBG.Rotate) & 63)) {
Opcode = SystemZ::RISBMux;
OpcodeVT = MVT::i32;
RISBG.Start &= 31;
RISBG.End &= 31;
}
SDValue Ops[5] = {
getUNDEF(DL, OpcodeVT),
convertTo(DL, OpcodeVT, RISBG.Input),
CurDAG->getTargetConstant(RISBG.Start, DL, MVT::i32),
CurDAG->getTargetConstant(RISBG.End | 128, DL, MVT::i32),
CurDAG->getTargetConstant(RISBG.Rotate, DL, MVT::i32)
};
SDValue New = convertTo(
DL, VT, SDValue(CurDAG->getMachineNode(Opcode, DL, OpcodeVT, Ops), 0));
ReplaceNode(N, New.getNode());
return true;
}
bool SystemZDAGToDAGISel::tryRxSBG(SDNode *N, unsigned Opcode) {
SDLoc DL(N);
EVT VT = N->getValueType(0);
if (!VT.isInteger() || VT.getSizeInBits() > 64)
return false;
// 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]))
// The widening or narrowing is expected to be free.
// Counting widening or narrowing as a saved operation will result in
// preferring an R*SBG over a simple shift/logical instruction.
if (RxSBG[I].Input.getOpcode() != ISD::ANY_EXTEND &&
RxSBG[I].Input.getOpcode() != ISD::TRUNCATE)
Count[I] += 1;
// Do nothing if neither operand is suitable.
if (Count[0] == 0 && Count[1] == 0)
return false;
// 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 (auto *Load = dyn_cast<LoadSDNode>(Op0.getNode()))
if (Load->getMemoryVT() == MVT::i8)
return false;
// 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;
// Prefer RISBGN if available, since it does not clobber CC.
if (Subtarget->hasMiscellaneousExtensions())
Opcode = SystemZ::RISBGN;
}
SDValue Ops[5] = {
convertTo(DL, MVT::i64, Op0),
convertTo(DL, MVT::i64, RxSBG[I].Input),
CurDAG->getTargetConstant(RxSBG[I].Start, DL, MVT::i32),
CurDAG->getTargetConstant(RxSBG[I].End, DL, MVT::i32),
CurDAG->getTargetConstant(RxSBG[I].Rotate, DL, MVT::i32)
};
SDValue New = convertTo(
DL, VT, SDValue(CurDAG->getMachineNode(Opcode, DL, MVT::i64, Ops), 0));
ReplaceNode(N, New.getNode());
return true;
}
void 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, DL, VT);
if (Op0.getNode())
Upper = CurDAG->getNode(Opcode, DL, VT, Op0, Upper);
{
// When we haven't passed in Op0, Upper will be a constant. In order to
// prevent folding back to the large immediate in `Or = getNode(...)` we run
// SelectCode first and end up with an opaque machine node. This means that
// we need to use a handle to keep track of Upper in case it gets CSE'd by
// SelectCode.
//
// Note that in the case where Op0 is passed in we could just call
// SelectCode(Upper) later, along with the SelectCode(Or), and avoid needing
// the handle at all, but it's fine to do it here.
//
// TODO: This is a pretty hacky way to do this. Can we do something that
// doesn't require a two paragraph explanation?
HandleSDNode Handle(Upper);
SelectCode(Upper.getNode());
Upper = Handle.getValue();
}
SDValue Lower = CurDAG->getConstant(LowerVal, DL, VT);
SDValue Or = CurDAG->getNode(Opcode, DL, VT, Upper, Lower);
ReplaceNode(Node, Or.getNode());
SelectCode(Or.getNode());
}
bool SystemZDAGToDAGISel::tryGather(SDNode *N, unsigned Opcode) {
SDValue ElemV = N->getOperand(2);
auto *ElemN = dyn_cast<ConstantSDNode>(ElemV);
if (!ElemN)
return false;
unsigned Elem = ElemN->getZExtValue();
EVT VT = N->getValueType(0);
if (Elem >= VT.getVectorNumElements())
return false;
auto *Load = dyn_cast<LoadSDNode>(N->getOperand(1));
if (!Load || !Load->hasOneUse())
return false;
if (Load->getMemoryVT().getSizeInBits() !=
Load->getValueType(0).getSizeInBits())
return false;
SDValue Base, Disp, Index;
if (!selectBDVAddr12Only(Load->getBasePtr(), ElemV, Base, Disp, Index) ||
Index.getValueType() != VT.changeVectorElementTypeToInteger())
return false;
SDLoc DL(Load);
SDValue Ops[] = {
N->getOperand(0), Base, Disp, Index,
CurDAG->getTargetConstant(Elem, DL, MVT::i32), Load->getChain()
};
SDNode *Res = CurDAG->getMachineNode(Opcode, DL, VT, MVT::Other, Ops);
ReplaceUses(SDValue(Load, 1), SDValue(Res, 1));
ReplaceNode(N, Res);
return true;
}
bool SystemZDAGToDAGISel::tryScatter(StoreSDNode *Store, unsigned Opcode) {
SDValue Value = Store->getValue();
if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
return false;
if (Store->getMemoryVT().getSizeInBits() != Value.getValueSizeInBits())
return false;
SDValue ElemV = Value.getOperand(1);
auto *ElemN = dyn_cast<ConstantSDNode>(ElemV);
if (!ElemN)
return false;
SDValue Vec = Value.getOperand(0);
EVT VT = Vec.getValueType();
unsigned Elem = ElemN->getZExtValue();
if (Elem >= VT.getVectorNumElements())
return false;
SDValue Base, Disp, Index;
if (!selectBDVAddr12Only(Store->getBasePtr(), ElemV, Base, Disp, Index) ||
Index.getValueType() != VT.changeVectorElementTypeToInteger())
return false;
SDLoc DL(Store);
SDValue Ops[] = {
Vec, Base, Disp, Index, CurDAG->getTargetConstant(Elem, DL, MVT::i32),
Store->getChain()
};
ReplaceNode(Store, CurDAG->getMachineNode(Opcode, DL, MVT::Other, Ops));
return true;
}
// Check whether or not the chain ending in StoreNode is suitable for doing
// the {load; op; store} to modify transformation.
static bool isFusableLoadOpStorePattern(StoreSDNode *StoreNode,
SDValue StoredVal, SelectionDAG *CurDAG,
LoadSDNode *&LoadNode,
SDValue &InputChain) {
// Is the stored value result 0 of the operation?
if (StoredVal.getResNo() != 0)
return false;
// Are there other uses of the loaded value than the operation?
if (!StoredVal.getNode()->hasNUsesOfValue(1, 0))
return false;
// Is the store non-extending and non-indexed?
if (!ISD::isNormalStore(StoreNode) || StoreNode->isNonTemporal())
return false;
SDValue Load = StoredVal->getOperand(0);
// Is the stored value a non-extending and non-indexed load?
if (!ISD::isNormalLoad(Load.getNode()))
return false;
// Return LoadNode by reference.
LoadNode = cast<LoadSDNode>(Load);
// Is store the only read of the loaded value?
if (!Load.hasOneUse())
return false;
// Is the address of the store the same as the load?
if (LoadNode->getBasePtr() != StoreNode->getBasePtr() ||
LoadNode->getOffset() != StoreNode->getOffset())
return false;
// Check if the chain is produced by the load or is a TokenFactor with
// the load output chain as an operand. Return InputChain by reference.
SDValue Chain = StoreNode->getChain();
bool ChainCheck = false;
if (Chain == Load.getValue(1)) {
ChainCheck = true;
InputChain = LoadNode->getChain();
} else if (Chain.getOpcode() == ISD::TokenFactor) {
SmallVector<SDValue, 4> ChainOps;
for (unsigned i = 0, e = Chain.getNumOperands(); i != e; ++i) {
SDValue Op = Chain.getOperand(i);
if (Op == Load.getValue(1)) {
ChainCheck = true;
// Drop Load, but keep its chain. No cycle check necessary.
ChainOps.push_back(Load.getOperand(0));
continue;
}
// Make sure using Op as part of the chain would not cause a cycle here.
// In theory, we could check whether the chain node is a predecessor of
// the load. But that can be very expensive. Instead visit the uses and
// make sure they all have smaller node id than the load.
int LoadId = LoadNode->getNodeId();
for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
UE = UI->use_end(); UI != UE; ++UI) {
if (UI.getUse().getResNo() != 0)
continue;
if (UI->getNodeId() > LoadId)
return false;
}
ChainOps.push_back(Op);
}
if (ChainCheck)
// Make a new TokenFactor with all the other input chains except
// for the load.
InputChain = CurDAG->getNode(ISD::TokenFactor, SDLoc(Chain),
MVT::Other, ChainOps);
}
if (!ChainCheck)
return false;
return true;
}
// Change a chain of {load; op; store} of the same value into a simple op
// through memory of that value, if the uses of the modified value and its
// address are suitable.
//
// The tablegen pattern memory operand pattern is currently not able to match
// the case where the CC on the original operation are used.
//
// See the equivalent routine in X86ISelDAGToDAG for further comments.
bool SystemZDAGToDAGISel::tryFoldLoadStoreIntoMemOperand(SDNode *Node) {
StoreSDNode *StoreNode = cast<StoreSDNode>(Node);
SDValue StoredVal = StoreNode->getOperand(1);
unsigned Opc = StoredVal->getOpcode();
SDLoc DL(StoreNode);
// Before we try to select anything, make sure this is memory operand size
// and opcode we can handle. Note that this must match the code below that
// actually lowers the opcodes.
EVT MemVT = StoreNode->getMemoryVT();
unsigned NewOpc = 0;
bool NegateOperand = false;
switch (Opc) {
default:
return false;
case SystemZISD::SSUBO:
NegateOperand = true;
LLVM_FALLTHROUGH;
case SystemZISD::SADDO:
if (MemVT == MVT::i32)
NewOpc = SystemZ::ASI;
else if (MemVT == MVT::i64)
NewOpc = SystemZ::AGSI;
else
return false;
break;
case SystemZISD::USUBO:
NegateOperand = true;
LLVM_FALLTHROUGH;
case SystemZISD::UADDO:
if (MemVT == MVT::i32)
NewOpc = SystemZ::ALSI;
else if (MemVT == MVT::i64)
NewOpc = SystemZ::ALGSI;
else
return false;
break;
}
LoadSDNode *LoadNode = nullptr;
SDValue InputChain;
if (!isFusableLoadOpStorePattern(StoreNode, StoredVal, CurDAG, LoadNode,
InputChain))
return false;
SDValue Operand = StoredVal.getOperand(1);
auto *OperandC = dyn_cast<ConstantSDNode>(Operand);
if (!OperandC)
return false;
auto OperandV = OperandC->getAPIntValue();
if (NegateOperand)
OperandV = -OperandV;
if (OperandV.getMinSignedBits() > 8)
return false;
Operand = CurDAG->getTargetConstant(OperandV, DL, MemVT);
SDValue Base, Disp;
if (!selectBDAddr20Only(StoreNode->getBasePtr(), Base, Disp))
return false;
SDValue Ops[] = { Base, Disp, Operand, InputChain };
MachineSDNode *Result =
CurDAG->getMachineNode(NewOpc, DL, MVT::i32, MVT::Other, Ops);
CurDAG->setNodeMemRefs(
Result, {StoreNode->getMemOperand(), LoadNode->getMemOperand()});
ReplaceUses(SDValue(StoreNode, 0), SDValue(Result, 1));
ReplaceUses(SDValue(StoredVal.getNode(), 1), SDValue(Result, 0));
CurDAG->RemoveDeadNode(Node);
return true;
}
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;
// 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() && Load->isDereferenceable())
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(MemoryLocation(V1, End1, Load->getAAInfo()),
MemoryLocation(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.
uint64_t Size = Load->getMemoryVT().getStoreSize();
if (Size > 1 && Size <= 8) {
// Prefer LHRL, LRL and LGRL.
if (SystemZISD::isPCREL(Load->getBasePtr().getOpcode()))
return false;
// Prefer STHRL, STRL and STGRL.
if (SystemZISD::isPCREL(Store->getBasePtr().getOpcode()))
return false;
}
return canUseBlockOperation(Store, Load);
}
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);
}
void SystemZDAGToDAGISel::Select(SDNode *Node) {
// If we have a custom node, we already have selected!
if (Node->isMachineOpcode()) {
LLVM_DEBUG(errs() << "== "; Node->dump(CurDAG); errs() << "\n");
Node->setNodeId(-1);
return;
}
unsigned Opcode = Node->getOpcode();
switch (Opcode) {
case ISD::OR:
if (Node->getOperand(1).getOpcode() != ISD::Constant)
if (tryRxSBG(Node, SystemZ::ROSBG))
return;
goto or_xor;
case ISD::XOR:
if (Node->getOperand(1).getOpcode() != ISD::Constant)
if (tryRxSBG(Node, SystemZ::RXSBG))
return;
// 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 both operands here happen to be
// constant, leave this to common code to optimize.
if (Node->getValueType(0) == MVT::i64 &&
Node->getOperand(0).getOpcode() != ISD::Constant)
if (auto *Op1 = dyn_cast<ConstantSDNode>(Node->getOperand(1))) {
uint64_t Val = Op1->getZExtValue();
if (!SystemZ::isImmLF(Val) && !SystemZ::isImmHF(Val)) {
splitLargeImmediate(Opcode, Node, Node->getOperand(0),
Val - uint32_t(Val), uint32_t(Val));
return;
}
}
break;
case ISD::AND:
if (Node->getOperand(1).getOpcode() != ISD::Constant)
if (tryRxSBG(Node, SystemZ::RNSBG))
return;
LLVM_FALLTHROUGH;
case ISD::ROTL:
case ISD::SHL:
case ISD::SRL:
case ISD::ZERO_EXTEND:
if (tryRISBGZero(Node))
return;
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)) {
splitLargeImmediate(ISD::OR, Node, SDValue(), Val - uint32_t(Val),
uint32_t(Val));
return;
}
}
break;
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. Likewise for constants in range for LOCHI.
if ((Op1.getOpcode() == ISD::LOAD && Op0.getOpcode() != ISD::LOAD) ||
(Subtarget->hasLoadStoreOnCond2() &&
Node->getValueType(0).isInteger() &&
Op1.getOpcode() == ISD::Constant &&
isInt<16>(cast<ConstantSDNode>(Op1)->getSExtValue()) &&
!(Op0.getOpcode() == ISD::Constant &&
isInt<16>(cast<ConstantSDNode>(Op0)->getSExtValue())))) {
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, SDLoc(Node),
CCMask.getValueType());
SDValue Op4 = Node->getOperand(4);
SDNode *UpdatedNode =
CurDAG->UpdateNodeOperands(Node, Op1, Op0, CCValid, CCMask, Op4);
if (UpdatedNode != Node) {
// In case this node already exists then replace Node with it.
ReplaceNode(Node, UpdatedNode);
Node = UpdatedNode;
}
}
break;
}
case ISD::INSERT_VECTOR_ELT: {
EVT VT = Node->getValueType(0);
unsigned ElemBitSize = VT.getScalarSizeInBits();
if (ElemBitSize == 32) {
if (tryGather(Node, SystemZ::VGEF))
return;
} else if (ElemBitSize == 64) {
if (tryGather(Node, SystemZ::VGEG))
return;
}
break;
}
case ISD::STORE: {
if (tryFoldLoadStoreIntoMemOperand(Node))
return;
auto *Store = cast<StoreSDNode>(Node);
unsigned ElemBitSize = Store->getValue().getValueSizeInBits();
if (ElemBitSize == 32) {
if (tryScatter(Store, SystemZ::VSCEF))
return;
} else if (ElemBitSize == 64) {
if (tryScatter(Store, SystemZ::VSCEG))
return;
}
break;
}
}
SelectCode(Node);
}
bool SystemZDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue &Op,
unsigned ConstraintID,
std::vector<SDValue> &OutOps) {
SystemZAddressingMode::AddrForm Form;
SystemZAddressingMode::DispRange DispRange;
SDValue Base, Disp, Index;
switch(ConstraintID) {
default:
llvm_unreachable("Unexpected asm memory constraint");
case InlineAsm::Constraint_i:
case InlineAsm::Constraint_Q:
// Accept an address with a short displacement, but no index.
Form = SystemZAddressingMode::FormBD;
DispRange = SystemZAddressingMode::Disp12Only;
break;
case InlineAsm::Constraint_R:
// Accept an address with a short displacement and an index.
Form = SystemZAddressingMode::FormBDXNormal;
DispRange = SystemZAddressingMode::Disp12Only;
break;
case InlineAsm::Constraint_S:
// Accept an address with a long displacement, but no index.
Form = SystemZAddressingMode::FormBD;
DispRange = SystemZAddressingMode::Disp20Only;
break;
case InlineAsm::Constraint_T:
case InlineAsm::Constraint_m:
case InlineAsm::Constraint_o:
// Accept an address with a long displacement and an index.
// m works the same as T, as this is the most general case.
// We don't really have any special handling of "offsettable"
// memory addresses, so just treat o the same as m.
Form = SystemZAddressingMode::FormBDXNormal;
DispRange = SystemZAddressingMode::Disp20Only;
break;
}
if (selectBDXAddr(Form, DispRange, Op, Base, Disp, Index)) {
const TargetRegisterClass *TRC =
Subtarget->getRegisterInfo()->getPointerRegClass(*MF);
SDLoc DL(Base);
SDValue RC = CurDAG->getTargetConstant(TRC->getID(), DL, MVT::i32);
// Make sure that the base address doesn't go into %r0.
// If it's a TargetFrameIndex or a fixed register, we shouldn't do anything.
if (Base.getOpcode() != ISD::TargetFrameIndex &&
Base.getOpcode() != ISD::Register) {
Base =
SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
DL, Base.getValueType(),
Base, RC), 0);
}
// Make sure that the index register isn't assigned to %r0 either.
if (Index.getOpcode() != ISD::Register) {
Index =
SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
DL, Index.getValueType(),
Index, RC), 0);
}
OutOps.push_back(Base);
OutOps.push_back(Disp);
OutOps.push_back(Index);
return false;
}
return true;
}
// IsProfitableToFold - Returns true if is profitable to fold the specific
// operand node N of U during instruction selection that starts at Root.
bool
SystemZDAGToDAGISel::IsProfitableToFold(SDValue N, SDNode *U,
SDNode *Root) const {
// We want to avoid folding a LOAD into an ICMP node if as a result
// we would be forced to spill the condition code into a GPR.
if (N.getOpcode() == ISD::LOAD && U->getOpcode() == SystemZISD::ICMP) {
if (!N.hasOneUse() || !U->hasOneUse())
return false;
// The user of the CC value will usually be a CopyToReg into the
// physical CC register, which in turn is glued and chained to the
// actual instruction that uses the CC value. Bail out if we have
// anything else than that.
SDNode *CCUser = *U->use_begin();
SDNode *CCRegUser = nullptr;
if (CCUser->getOpcode() == ISD::CopyToReg ||
cast<RegisterSDNode>(CCUser->getOperand(1))->getReg() == SystemZ::CC) {
for (auto *U : CCUser->uses()) {
if (CCRegUser == nullptr)
CCRegUser = U;
else if (CCRegUser != U)
return false;
}
}
if (CCRegUser == nullptr)
return false;
// If the actual instruction is a branch, the only thing that remains to be
// checked is whether the CCUser chain is a predecessor of the load.
if (CCRegUser->isMachineOpcode() &&
CCRegUser->getMachineOpcode() == SystemZ::BRC)
return !N->isPredecessorOf(CCUser->getOperand(0).getNode());
// Otherwise, the instruction may have multiple operands, and we need to
// verify that none of them are a predecessor of the load. This is exactly
// the same check that would be done by common code if the CC setter were
// glued to the CC user, so simply invoke that check here.
if (!IsLegalToFold(N, U, CCRegUser, OptLevel, false))
return false;
}
return true;
}
namespace {
// Represents a sequence for extracting a 0/1 value from an IPM result:
// (((X ^ XORValue) + AddValue) >> Bit)
struct IPMConversion {
IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
: XORValue(xorValue), AddValue(addValue), Bit(bit) {}
int64_t XORValue;
int64_t AddValue;
unsigned Bit;
};
} // end anonymous namespace
// Return a sequence for getting a 1 from an IPM result when CC has a
// value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
// The handling of CC values outside CCValid doesn't matter.
static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
// Deal with cases where the result can be taken directly from a bit
// of the IPM result.
if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
return IPMConversion(0, 0, SystemZ::IPM_CC);
if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
// Deal with cases where we can add a value to force the sign bit
// to contain the right value. Putting the bit in 31 means we can
// use SRL rather than RISBG(L), and also makes it easier to get a
// 0/-1 value, so it has priority over the other tests below.
//
// These sequences rely on the fact that the upper two bits of the
// IPM result are zero.
uint64_t TopBit = uint64_t(1) << 31;
if (CCMask == (CCValid & SystemZ::CCMASK_0))
return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0
| SystemZ::CCMASK_1
| SystemZ::CCMASK_2)))
return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & SystemZ::CCMASK_3))
return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_1
| SystemZ::CCMASK_2
| SystemZ::CCMASK_3)))
return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
// Next try inverting the value and testing a bit. 0/1 could be
// handled this way too, but we dealt with that case above.
if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
return IPMConversion(-1, 0, SystemZ::IPM_CC);
// Handle cases where adding a value forces a non-sign bit to contain
// the right value.
if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
// The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
// can be done by inverting the low CC bit and applying one of the
// sign-based extractions above.
if (CCMask == (CCValid & SystemZ::CCMASK_1))
return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & SystemZ::CCMASK_2))
return IPMConversion(1 << SystemZ::IPM_CC,
TopBit - (3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0
| SystemZ::CCMASK_1
| SystemZ::CCMASK_3)))
return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
if (CCMask == (CCValid & (SystemZ::CCMASK_0
| SystemZ::CCMASK_2
| SystemZ::CCMASK_3)))
return IPMConversion(1 << SystemZ::IPM_CC,
TopBit - (1 << SystemZ::IPM_CC), 31);
llvm_unreachable("Unexpected CC combination");
}
SDValue SystemZDAGToDAGISel::expandSelectBoolean(SDNode *Node) {
auto *TrueOp = dyn_cast<ConstantSDNode>(Node->getOperand(0));
auto *FalseOp = dyn_cast<ConstantSDNode>(Node->getOperand(1));
if (!TrueOp || !FalseOp)
return SDValue();
if (FalseOp->getZExtValue() != 0)
return SDValue();
if (TrueOp->getSExtValue() != 1 && TrueOp->getSExtValue() != -1)
return SDValue();
auto *CCValidOp = dyn_cast<ConstantSDNode>(Node->getOperand(2));
auto *CCMaskOp = dyn_cast<ConstantSDNode>(Node->getOperand(3));
if (!CCValidOp || !CCMaskOp)
return SDValue();
int CCValid = CCValidOp->getZExtValue();
int CCMask = CCMaskOp->getZExtValue();
SDLoc DL(Node);
SDValue CCReg = Node->getOperand(4);
IPMConversion IPM = getIPMConversion(CCValid, CCMask);
SDValue Result = CurDAG->getNode(SystemZISD::IPM, DL, MVT::i32, CCReg);
if (IPM.XORValue)
Result = CurDAG->getNode(ISD::XOR, DL, MVT::i32, Result,
CurDAG->getConstant(IPM.XORValue, DL, MVT::i32));
if (IPM.AddValue)
Result = CurDAG->getNode(ISD::ADD, DL, MVT::i32, Result,
CurDAG->getConstant(IPM.AddValue, DL, MVT::i32));
EVT VT = Node->getValueType(0);
if (VT == MVT::i32 && IPM.Bit == 31) {
unsigned ShiftOp = TrueOp->getSExtValue() == 1 ? ISD::SRL : ISD::SRA;
Result = CurDAG->getNode(ShiftOp, DL, MVT::i32, Result,
CurDAG->getConstant(IPM.Bit, DL, MVT::i32));
} else {
if (VT != MVT::i32)
Result = CurDAG->getNode(ISD::ANY_EXTEND, DL, VT, Result);
if (TrueOp->getSExtValue() == 1) {
// The SHR/AND sequence should get optimized to an RISBG.
Result = CurDAG->getNode(ISD::SRL, DL, VT, Result,
CurDAG->getConstant(IPM.Bit, DL, MVT::i32));
Result = CurDAG->getNode(ISD::AND, DL, VT, Result,
CurDAG->getConstant(1, DL, VT));
} else {
// Sign-extend from IPM.Bit using a pair of shifts.
int ShlAmt = VT.getSizeInBits() - 1 - IPM.Bit;
int SraAmt = VT.getSizeInBits() - 1;
Result = CurDAG->getNode(ISD::SHL, DL, VT, Result,
CurDAG->getConstant(ShlAmt, DL, MVT::i32));
Result = CurDAG->getNode(ISD::SRA, DL, VT, Result,
CurDAG->getConstant(SraAmt, DL, MVT::i32));
}
}
return Result;
}
void SystemZDAGToDAGISel::PreprocessISelDAG() {
// If we have conditional immediate loads, we always prefer
// using those over an IPM sequence.
if (Subtarget->hasLoadStoreOnCond2())
return;
bool MadeChange = false;
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
E = CurDAG->allnodes_end();
I != E;) {
SDNode *N = &*I++;
if (N->use_empty())
continue;
SDValue Res;
switch (N->getOpcode()) {
default: break;
case SystemZISD::SELECT_CCMASK:
Res = expandSelectBoolean(N);
break;
}
if (Res) {
LLVM_DEBUG(dbgs() << "SystemZ DAG preprocessing replacing:\nOld: ");
LLVM_DEBUG(N->dump(CurDAG));
LLVM_DEBUG(dbgs() << "\nNew: ");
LLVM_DEBUG(Res.getNode()->dump(CurDAG));
LLVM_DEBUG(dbgs() << "\n");
CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res);
MadeChange = true;
}
}
if (MadeChange)
CurDAG->RemoveDeadNodes();
}
Reference in New Issue View Git Blame Copy Permalink