#include "SPUHazardRecognizers.h"
#include "SPUFrameInfo.h"
#include "SPURegisterNames.h"
+#include "SPUTargetMachine.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Constants.h"
#include "llvm/GlobalValue.h"
#include "llvm/Intrinsics.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/Compiler.h"
+#include "llvm/Support/raw_ostream.h"
using namespace llvm;
return isS10Constant(CN->getSExtValue());
}
-#if 0
- //! SDNode predicate for sign-extended, 10-bit immediate values
- bool
- isI32IntS10Immediate(SDNode *N)
- {
- return (N->getOpcode() == ISD::Constant
- && isI32IntS10Immediate(cast<ConstantSDNode>(N)));
- }
-#endif
-
//! ConstantSDNode predicate for i32 unsigned 10-bit immediate values
bool
isI32IntU10Immediate(ConstantSDNode *CN)
bool
isI16IntS10Immediate(SDNode *N)
{
- return (N->getOpcode() == ISD::Constant
- && isI16IntS10Immediate(cast<ConstantSDNode>(N)));
+ ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
+ return (CN != 0 && isI16IntS10Immediate(CN));
}
//! ConstantSDNode predicate for i16 unsigned 10-bit immediate values
bool
isIntS16Immediate(ConstantSDNode *CN, short &Imm)
{
- MVT vt = CN->getValueType(0);
+ EVT vt = CN->getValueType(0);
Imm = (short) CN->getZExtValue();
if (vt.getSimpleVT() >= MVT::i1 && vt.getSimpleVT() <= MVT::i16) {
return true;
static bool
isFPS16Immediate(ConstantFPSDNode *FPN, short &Imm)
{
- MVT vt = FPN->getValueType(0);
+ EVT vt = FPN->getValueType(0);
if (vt == MVT::f32) {
int val = FloatToBits(FPN->getValueAPF().convertToFloat());
int sval = (int) ((val << 16) >> 16);
}
bool
- isHighLow(const SDValue &Op)
+ isHighLow(const SDValue &Op)
{
return (Op.getOpcode() == SPUISD::IndirectAddr
&& ((Op.getOperand(0).getOpcode() == SPUISD::Hi
}
//===------------------------------------------------------------------===//
- //! MVT to "useful stuff" mapping structure:
+ //! EVT to "useful stuff" mapping structure:
struct valtype_map_s {
- MVT VT;
+ EVT VT;
unsigned ldresult_ins; /// LDRESULT instruction (0 = undefined)
bool ldresult_imm; /// LDRESULT instruction requires immediate?
- int prefslot_byte; /// Byte offset of the "preferred" slot
+ unsigned lrinst; /// LR instruction
};
const valtype_map_s valtype_map[] = {
- { MVT::i1, 0, false, 3 },
- { MVT::i8, SPU::ORBIr8, true, 3 },
- { MVT::i16, SPU::ORHIr16, true, 2 },
- { MVT::i32, SPU::ORIr32, true, 0 },
- { MVT::i64, SPU::ORr64, false, 0 },
- { MVT::f32, SPU::ORf32, false, 0 },
- { MVT::f64, SPU::ORf64, false, 0 },
+ { MVT::i8, SPU::ORBIr8, true, SPU::LRr8 },
+ { MVT::i16, SPU::ORHIr16, true, SPU::LRr16 },
+ { MVT::i32, SPU::ORIr32, true, SPU::LRr32 },
+ { MVT::i64, SPU::ORr64, false, SPU::LRr64 },
+ { MVT::f32, SPU::ORf32, false, SPU::LRf32 },
+ { MVT::f64, SPU::ORf64, false, SPU::LRf64 },
// vector types... (sigh!)
- { MVT::v16i8, 0, false, 0 },
- { MVT::v8i16, 0, false, 0 },
- { MVT::v4i32, 0, false, 0 },
- { MVT::v2i64, 0, false, 0 },
- { MVT::v4f32, 0, false, 0 },
- { MVT::v2f64, 0, false, 0 }
+ { MVT::v16i8, 0, false, SPU::LRv16i8 },
+ { MVT::v8i16, 0, false, SPU::LRv8i16 },
+ { MVT::v4i32, 0, false, SPU::LRv4i32 },
+ { MVT::v2i64, 0, false, SPU::LRv2i64 },
+ { MVT::v4f32, 0, false, SPU::LRv4f32 },
+ { MVT::v2f64, 0, false, SPU::LRv2f64 }
};
const size_t n_valtype_map = sizeof(valtype_map) / sizeof(valtype_map[0]);
- const valtype_map_s *getValueTypeMapEntry(MVT VT)
+ const valtype_map_s *getValueTypeMapEntry(EVT VT)
{
const valtype_map_s *retval = 0;
for (size_t i = 0; i < n_valtype_map; ++i) {
#ifndef NDEBUG
if (retval == 0) {
- cerr << "SPUISelDAGToDAG.cpp: getValueTypeMapEntry returns NULL for "
- << VT.getMVTString()
- << "\n";
- abort();
+ std::string msg;
+ raw_string_ostream Msg(msg);
+ Msg << "SPUISelDAGToDAG.cpp: getValueTypeMapEntry returns NULL for "
+ << VT.getEVTString();
+ llvm_report_error(Msg.str());
}
#endif
return retval;
}
-}
-namespace {
+ //! Generate the carry-generate shuffle mask.
+ SDValue getCarryGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
+ SmallVector<SDValue, 16 > ShufBytes;
-//===--------------------------------------------------------------------===//
-/// SPUDAGToDAGISel - Cell SPU-specific code to select SPU machine
-/// instructions for SelectionDAG operations.
-///
-class SPUDAGToDAGISel :
- public SelectionDAGISel
-{
- SPUTargetMachine &TM;
- SPUTargetLowering &SPUtli;
- unsigned GlobalBaseReg;
-
-public:
- explicit SPUDAGToDAGISel(SPUTargetMachine &tm) :
- SelectionDAGISel(*tm.getTargetLowering()),
- TM(tm),
- SPUtli(*tm.getTargetLowering())
- {}
-
- virtual bool runOnFunction(Function &Fn) {
- // Make sure we re-emit a set of the global base reg if necessary
- GlobalBaseReg = 0;
- SelectionDAGISel::runOnFunction(Fn);
- return true;
- }
-
- /// getI32Imm - Return a target constant with the specified value, of type
- /// i32.
- inline SDValue getI32Imm(uint32_t Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i32);
- }
+ // Create the shuffle mask for "rotating" the borrow up one register slot
+ // once the borrow is generated.
+ ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
+ ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
+ ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
+ ShufBytes.push_back(DAG.getConstant(0x80808080, MVT::i32));
- /// getI64Imm - Return a target constant with the specified value, of type
- /// i64.
- inline SDValue getI64Imm(uint64_t Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i64);
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
+ &ShufBytes[0], ShufBytes.size());
}
-
- /// getSmallIPtrImm - Return a target constant of pointer type.
- inline SDValue getSmallIPtrImm(unsigned Imm) {
- return CurDAG->getTargetConstant(Imm, SPUtli.getPointerTy());
+
+ //! Generate the borrow-generate shuffle mask
+ SDValue getBorrowGenerateShufMask(SelectionDAG &DAG, DebugLoc dl) {
+ SmallVector<SDValue, 16 > ShufBytes;
+
+ // Create the shuffle mask for "rotating" the borrow up one register slot
+ // once the borrow is generated.
+ ShufBytes.push_back(DAG.getConstant(0x04050607, MVT::i32));
+ ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
+ ShufBytes.push_back(DAG.getConstant(0x0c0d0e0f, MVT::i32));
+ ShufBytes.push_back(DAG.getConstant(0xc0c0c0c0, MVT::i32));
+
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
+ &ShufBytes[0], ShufBytes.size());
}
- /// Select - Convert the specified operand from a target-independent to a
- /// target-specific node if it hasn't already been changed.
- SDNode *Select(SDValue Op);
-
- //! Returns true if the address N is an A-form (local store) address
- bool SelectAFormAddr(SDValue Op, SDValue N, SDValue &Base,
- SDValue &Index);
-
- //! D-form address predicate
- bool SelectDFormAddr(SDValue Op, SDValue N, SDValue &Base,
- SDValue &Index);
-
- /// Alternate D-form address using i7 offset predicate
- bool SelectDForm2Addr(SDValue Op, SDValue N, SDValue &Disp,
- SDValue &Base);
-
- /// D-form address selection workhorse
- bool DFormAddressPredicate(SDValue Op, SDValue N, SDValue &Disp,
- SDValue &Base, int minOffset, int maxOffset);
-
- //! Address predicate if N can be expressed as an indexed [r+r] operation.
- bool SelectXFormAddr(SDValue Op, SDValue N, SDValue &Base,
- SDValue &Index);
-
- /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
- /// inline asm expressions.
- virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
- char ConstraintCode,
- std::vector<SDValue> &OutOps) {
- SDValue Op0, Op1;
- switch (ConstraintCode) {
- default: return true;
- case 'm': // memory
- if (!SelectDFormAddr(Op, Op, Op0, Op1)
- && !SelectAFormAddr(Op, Op, Op0, Op1))
- SelectXFormAddr(Op, Op, Op0, Op1);
- break;
- case 'o': // offsetable
- if (!SelectDFormAddr(Op, Op, Op0, Op1)
- && !SelectAFormAddr(Op, Op, Op0, Op1)) {
- Op0 = Op;
- Op1 = getSmallIPtrImm(0);
+ //===------------------------------------------------------------------===//
+ /// SPUDAGToDAGISel - Cell SPU-specific code to select SPU machine
+ /// instructions for SelectionDAG operations.
+ ///
+ class SPUDAGToDAGISel :
+ public SelectionDAGISel
+ {
+ SPUTargetMachine &TM;
+ SPUTargetLowering &SPUtli;
+ unsigned GlobalBaseReg;
+
+ public:
+ explicit SPUDAGToDAGISel(SPUTargetMachine &tm) :
+ SelectionDAGISel(tm),
+ TM(tm),
+ SPUtli(*tm.getTargetLowering())
+ { }
+
+ virtual bool runOnMachineFunction(MachineFunction &MF) {
+ // Make sure we re-emit a set of the global base reg if necessary
+ GlobalBaseReg = 0;
+ SelectionDAGISel::runOnMachineFunction(MF);
+ return true;
+ }
+
+ /// getI32Imm - Return a target constant with the specified value, of type
+ /// i32.
+ inline SDValue getI32Imm(uint32_t Imm) {
+ return CurDAG->getTargetConstant(Imm, MVT::i32);
+ }
+
+ /// getI64Imm - Return a target constant with the specified value, of type
+ /// i64.
+ inline SDValue getI64Imm(uint64_t Imm) {
+ return CurDAG->getTargetConstant(Imm, MVT::i64);
+ }
+
+ /// getSmallIPtrImm - Return a target constant of pointer type.
+ inline SDValue getSmallIPtrImm(unsigned Imm) {
+ return CurDAG->getTargetConstant(Imm, SPUtli.getPointerTy());
}
- break;
- case 'v': // not offsetable
+
+ SDNode *emitBuildVector(SDValue build_vec) {
+ EVT vecVT = build_vec.getValueType();
+ EVT eltVT = vecVT.getVectorElementType();
+ SDNode *bvNode = build_vec.getNode();
+ DebugLoc dl = bvNode->getDebugLoc();
+
+ // Check to see if this vector can be represented as a CellSPU immediate
+ // constant by invoking all of the instruction selection predicates:
+ if (((vecVT == MVT::v8i16) &&
+ (SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i16).getNode() != 0)) ||
+ ((vecVT == MVT::v4i32) &&
+ ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
+ (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
+ (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i32).getNode() != 0) ||
+ (SPU::get_v4i32_imm(bvNode, *CurDAG).getNode() != 0))) ||
+ ((vecVT == MVT::v2i64) &&
+ ((SPU::get_vec_i16imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
+ (SPU::get_ILHUvec_imm(bvNode, *CurDAG, MVT::i64).getNode() != 0) ||
+ (SPU::get_vec_u18imm(bvNode, *CurDAG, MVT::i64).getNode() != 0))))
+ return Select(build_vec);
+
+ // No, need to emit a constant pool spill:
+ std::vector<Constant*> CV;
+
+ for (size_t i = 0; i < build_vec.getNumOperands(); ++i) {
+ ConstantSDNode *V = dyn_cast<ConstantSDNode > (build_vec.getOperand(i));
+ CV.push_back(const_cast<ConstantInt *> (V->getConstantIntValue()));
+ }
+
+ Constant *CP = ConstantVector::get(CV);
+ SDValue CPIdx = CurDAG->getConstantPool(CP, SPUtli.getPointerTy());
+ unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
+ SDValue CGPoolOffset =
+ SPU::LowerConstantPool(CPIdx, *CurDAG,
+ SPUtli.getSPUTargetMachine());
+ return SelectCode(CurDAG->getLoad(build_vec.getValueType(), dl,
+ CurDAG->getEntryNode(), CGPoolOffset,
+ PseudoSourceValue::getConstantPool(), 0,
+ false, Alignment));
+ }
+
+ /// Select - Convert the specified operand from a target-independent to a
+ /// target-specific node if it hasn't already been changed.
+ SDNode *Select(SDValue Op);
+
+ //! Emit the instruction sequence for i64 shl
+ SDNode *SelectSHLi64(SDValue &Op, EVT OpVT);
+
+ //! Emit the instruction sequence for i64 srl
+ SDNode *SelectSRLi64(SDValue &Op, EVT OpVT);
+
+ //! Emit the instruction sequence for i64 sra
+ SDNode *SelectSRAi64(SDValue &Op, EVT OpVT);
+
+ //! Emit the necessary sequence for loading i64 constants:
+ SDNode *SelectI64Constant(SDValue &Op, EVT OpVT, DebugLoc dl);
+
+ //! Alternate instruction emit sequence for loading i64 constants
+ SDNode *SelectI64Constant(uint64_t i64const, EVT OpVT, DebugLoc dl);
+
+ //! Returns true if the address N is an A-form (local store) address
+ bool SelectAFormAddr(SDValue Op, SDValue N, SDValue &Base,
+ SDValue &Index);
+
+ //! D-form address predicate
+ bool SelectDFormAddr(SDValue Op, SDValue N, SDValue &Base,
+ SDValue &Index);
+
+ /// Alternate D-form address using i7 offset predicate
+ bool SelectDForm2Addr(SDValue Op, SDValue N, SDValue &Disp,
+ SDValue &Base);
+
+ /// D-form address selection workhorse
+ bool DFormAddressPredicate(SDValue Op, SDValue N, SDValue &Disp,
+ SDValue &Base, int minOffset, int maxOffset);
+
+ //! Address predicate if N can be expressed as an indexed [r+r] operation.
+ bool SelectXFormAddr(SDValue Op, SDValue N, SDValue &Base,
+ SDValue &Index);
+
+ /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
+ /// inline asm expressions.
+ virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
+ char ConstraintCode,
+ std::vector<SDValue> &OutOps) {
+ SDValue Op0, Op1;
+ switch (ConstraintCode) {
+ default: return true;
+ case 'm': // memory
+ if (!SelectDFormAddr(Op, Op, Op0, Op1)
+ && !SelectAFormAddr(Op, Op, Op0, Op1))
+ SelectXFormAddr(Op, Op, Op0, Op1);
+ break;
+ case 'o': // offsetable
+ if (!SelectDFormAddr(Op, Op, Op0, Op1)
+ && !SelectAFormAddr(Op, Op, Op0, Op1)) {
+ Op0 = Op;
+ Op1 = getSmallIPtrImm(0);
+ }
+ break;
+ case 'v': // not offsetable
#if 1
- assert(0 && "InlineAsmMemoryOperand 'v' constraint not handled.");
+ llvm_unreachable("InlineAsmMemoryOperand 'v' constraint not handled.");
#else
- SelectAddrIdxOnly(Op, Op, Op0, Op1);
+ SelectAddrIdxOnly(Op, Op, Op0, Op1);
#endif
- break;
+ break;
+ }
+
+ OutOps.push_back(Op0);
+ OutOps.push_back(Op1);
+ return false;
}
-
- OutOps.push_back(Op0);
- OutOps.push_back(Op1);
- return false;
- }
- /// InstructionSelect - This callback is invoked by
- /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
- virtual void InstructionSelect();
-
- virtual const char *getPassName() const {
- return "Cell SPU DAG->DAG Pattern Instruction Selection";
- }
-
- /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
- /// this target when scheduling the DAG.
- virtual HazardRecognizer *CreateTargetHazardRecognizer() {
- const TargetInstrInfo *II = TM.getInstrInfo();
- assert(II && "No InstrInfo?");
- return new SPUHazardRecognizer(*II);
- }
+ /// InstructionSelect - This callback is invoked by
+ /// SelectionDAGISel when it has created a SelectionDAG for us to codegen.
+ virtual void InstructionSelect();
- // Include the pieces autogenerated from the target description.
-#include "SPUGenDAGISel.inc"
-};
+ virtual const char *getPassName() const {
+ return "Cell SPU DAG->DAG Pattern Instruction Selection";
+ }
+
+ /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
+ /// this target when scheduling the DAG.
+ virtual ScheduleHazardRecognizer *CreateTargetHazardRecognizer() {
+ const TargetInstrInfo *II = TM.getInstrInfo();
+ assert(II && "No InstrInfo?");
+ return new SPUHazardRecognizer(*II);
+ }
+ // Include the pieces autogenerated from the target description.
+#include "SPUGenDAGISel.inc"
+ };
}
/// InstructionSelect - This callback is invoked by
}
/*!
- \arg Op The ISD instructio operand
+ \arg Op The ISD instruction operand
\arg N The address to be tested
\arg Base The base address
\arg Index The base address index
SPUDAGToDAGISel::SelectAFormAddr(SDValue Op, SDValue N, SDValue &Base,
SDValue &Index) {
// These match the addr256k operand type:
- MVT OffsVT = MVT::i16;
+ EVT OffsVT = MVT::i16;
SDValue Zero = CurDAG->getTargetConstant(0, OffsVT);
switch (N.getOpcode()) {
case ISD::Constant:
case ISD::ConstantPool:
case ISD::GlobalAddress:
- cerr << "SPU SelectAFormAddr: Constant/Pool/Global not lowered.\n";
- abort();
+ llvm_report_error("SPU SelectAFormAddr: Constant/Pool/Global not lowered.");
/*NOTREACHED*/
case ISD::TargetConstant:
case ISD::TargetGlobalAddress:
case ISD::TargetJumpTable:
- cerr << "SPUSelectAFormAddr: Target Constant/Pool/Global not wrapped as "
- << "A-form address.\n";
- abort();
+ llvm_report_error("SPUSelectAFormAddr: Target Constant/Pool/Global "
+ "not wrapped as A-form address.");
/*NOTREACHED*/
- case SPUISD::AFormAddr:
+ case SPUISD::AFormAddr:
// Just load from memory if there's only a single use of the location,
// otherwise, this will get handled below with D-form offset addresses
if (N.hasOneUse()) {
return false;
}
-bool
+bool
SPUDAGToDAGISel::SelectDForm2Addr(SDValue Op, SDValue N, SDValue &Disp,
SDValue &Base) {
const int minDForm2Offset = -(1 << 7);
SDValue &Index, int minOffset,
int maxOffset) {
unsigned Opc = N.getOpcode();
- MVT PtrTy = SPUtli.getPointerTy();
+ EVT PtrTy = SPUtli.getPointerTy();
if (Opc == ISD::FrameIndex) {
// Stack frame index must be less than 512 (divided by 16):
FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(N);
int FI = int(FIN->getIndex());
- DEBUG(cerr << "SelectDFormAddr: ISD::FrameIndex = "
+ DEBUG(errs() << "SelectDFormAddr: ISD::FrameIndex = "
<< FI << "\n");
if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
Base = CurDAG->getTargetConstant(0, PtrTy);
if (Op0.getOpcode() == ISD::FrameIndex) {
FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op0);
int FI = int(FIN->getIndex());
- DEBUG(cerr << "SelectDFormAddr: ISD::ADD offset = " << offset
+ DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
<< " frame index = " << FI << "\n");
if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
if (Op1.getOpcode() == ISD::FrameIndex) {
FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op1);
int FI = int(FIN->getIndex());
- DEBUG(cerr << "SelectDFormAddr: ISD::ADD offset = " << offset
+ DEBUG(errs() << "SelectDFormAddr: ISD::ADD offset = " << offset
<< " frame index = " << FI << "\n");
if (SPUFrameInfo::FItoStackOffset(FI) < maxOffset) {
ConstantSDNode *CN = cast<ConstantSDNode>(Op0);
offset = int32_t(CN->getSExtValue());
idxOp = Op1;
- }
+ }
if (offset >= minOffset && offset <= maxOffset) {
Base = CurDAG->getTargetConstant(offset, PtrTy);
else
Addr = N; // Register
- if (OpOpc == ISD::STORE)
- Offs = Op.getOperand(3);
- else
- Offs = Op.getOperand(2); // LOAD
+ Offs = ((OpOpc == ISD::STORE) ? Op.getOperand(3) : Op.getOperand(2));
if (Offs.getOpcode() == ISD::Constant || Offs.getOpcode() == ISD::UNDEF) {
if (Offs.getOpcode() == ISD::UNDEF)
Index = Addr;
return true;
}
+ } else {
+ /* If otherwise unadorned, default to D-form address with 0 offset: */
+ if (Opc == ISD::CopyFromReg) {
+ Index = N.getOperand(1);
+ } else {
+ Index = N;
+ }
+
+ Base = CurDAG->getTargetConstant(0, Index.getValueType());
+ return true;
}
}
SDValue &Index) {
if (!SelectAFormAddr(Op, N, Base, Index)
&& !SelectDFormAddr(Op, N, Base, Index)) {
- // default form of a X-form address is r(r) in operands 0 and 1:
- SDValue Op0 = N.getOperand(0);
- SDValue Op1 = N.getOperand(1);
-
- if (Op0.getOpcode() == ISD::Register && Op1.getOpcode() == ISD::Register) {
- Base = Op0;
- Index = Op1;
- return true;
- }
+ // If the address is neither A-form or D-form, punt and use an X-form
+ // address:
+ Base = N.getOperand(1);
+ Index = N.getOperand(0);
+ return true;
}
return false;
unsigned Opc = N->getOpcode();
int n_ops = -1;
unsigned NewOpc;
- MVT OpVT = Op.getValueType();
+ EVT OpVT = Op.getValueType();
SDValue Ops[8];
+ DebugLoc dl = N->getDebugLoc();
if (N->isMachineOpcode()) {
return NULL; // Already selected.
- } else if (Opc == ISD::FrameIndex) {
- // Selects to (add $sp, FI * stackSlotSize)
- int FI =
- SPUFrameInfo::FItoStackOffset(cast<FrameIndexSDNode>(N)->getIndex());
- MVT PtrVT = SPUtli.getPointerTy();
-
- // Adjust stack slot to actual offset in frame:
- if (isS10Constant(FI)) {
- DEBUG(cerr << "SPUDAGToDAGISel: Replacing FrameIndex with AIr32 $sp, "
- << FI
- << "\n");
+ }
+
+ if (Opc == ISD::FrameIndex) {
+ int FI = cast<FrameIndexSDNode>(N)->getIndex();
+ SDValue TFI = CurDAG->getTargetFrameIndex(FI, Op.getValueType());
+ SDValue Imm0 = CurDAG->getTargetConstant(0, Op.getValueType());
+
+ if (FI < 128) {
NewOpc = SPU::AIr32;
- Ops[0] = CurDAG->getRegister(SPU::R1, PtrVT);
- Ops[1] = CurDAG->getTargetConstant(FI, PtrVT);
+ Ops[0] = TFI;
+ Ops[1] = Imm0;
n_ops = 2;
} else {
- DEBUG(cerr << "SPUDAGToDAGISel: Replacing FrameIndex with Ar32 $sp, "
- << FI
- << "\n");
NewOpc = SPU::Ar32;
- Ops[0] = CurDAG->getRegister(SPU::R1, PtrVT);
- Ops[1] = CurDAG->getConstant(FI, PtrVT);
+ Ops[0] = CurDAG->getRegister(SPU::R1, Op.getValueType());
+ Ops[1] = SDValue(CurDAG->getTargetNode(SPU::ILAr32, dl, Op.getValueType(),
+ TFI, Imm0), 0);
n_ops = 2;
}
- } else if (Opc == ISD::ZERO_EXTEND) {
- // (zero_extend:i16 (and:i8 <arg>, <const>))
- const SDValue &Op1 = N->getOperand(0);
-
- if (Op.getValueType() == MVT::i16 && Op1.getValueType() == MVT::i8) {
- if (Op1.getOpcode() == ISD::AND) {
- // Fold this into a single ANDHI. This is often seen in expansions of i1
- // to i8, then i8 to i16 in logical/branching operations.
- DEBUG(cerr << "CellSPU: Coalescing (zero_extend:i16 (and:i8 "
- "<arg>, <const>))\n");
- NewOpc = SPU::ANDHIi8i16;
- Ops[0] = Op1.getOperand(0);
- Ops[1] = Op1.getOperand(1);
- n_ops = 2;
+ } else if (Opc == ISD::Constant && OpVT == MVT::i64) {
+ // Catch the i64 constants that end up here. Note: The backend doesn't
+ // attempt to legalize the constant (it's useless because DAGCombiner
+ // will insert 64-bit constants and we can't stop it).
+ return SelectI64Constant(Op, OpVT, Op.getDebugLoc());
+ } else if ((Opc == ISD::ZERO_EXTEND || Opc == ISD::ANY_EXTEND)
+ && OpVT == MVT::i64) {
+ SDValue Op0 = Op.getOperand(0);
+ EVT Op0VT = Op0.getValueType();
+ EVT Op0VecVT = EVT::getVectorVT(*CurDAG->getContext(),
+ Op0VT, (128 / Op0VT.getSizeInBits()));
+ EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(),
+ OpVT, (128 / OpVT.getSizeInBits()));
+ SDValue shufMask;
+
+ switch (Op0VT.getSimpleVT().SimpleTy) {
+ default:
+ llvm_report_error("CellSPU Select: Unhandled zero/any extend EVT");
+ /*NOTREACHED*/
+ case MVT::i32:
+ shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
+ CurDAG->getConstant(0x80808080, MVT::i32),
+ CurDAG->getConstant(0x00010203, MVT::i32),
+ CurDAG->getConstant(0x80808080, MVT::i32),
+ CurDAG->getConstant(0x08090a0b, MVT::i32));
+ break;
+
+ case MVT::i16:
+ shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
+ CurDAG->getConstant(0x80808080, MVT::i32),
+ CurDAG->getConstant(0x80800203, MVT::i32),
+ CurDAG->getConstant(0x80808080, MVT::i32),
+ CurDAG->getConstant(0x80800a0b, MVT::i32));
+ break;
+
+ case MVT::i8:
+ shufMask = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
+ CurDAG->getConstant(0x80808080, MVT::i32),
+ CurDAG->getConstant(0x80808003, MVT::i32),
+ CurDAG->getConstant(0x80808080, MVT::i32),
+ CurDAG->getConstant(0x8080800b, MVT::i32));
+ break;
+ }
+
+ SDNode *shufMaskLoad = emitBuildVector(shufMask);
+ SDNode *PromoteScalar =
+ SelectCode(CurDAG->getNode(SPUISD::PREFSLOT2VEC, dl, Op0VecVT, Op0));
+
+ SDValue zextShuffle =
+ CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
+ SDValue(PromoteScalar, 0),
+ SDValue(PromoteScalar, 0),
+ SDValue(shufMaskLoad, 0));
+
+ // N.B.: BIT_CONVERT replaces and updates the zextShuffle node, so we
+ // re-use it in the VEC2PREFSLOT selection without needing to explicitly
+ // call SelectCode (it's already done for us.)
+ SelectCode(CurDAG->getNode(ISD::BIT_CONVERT, dl, OpVecVT, zextShuffle));
+ return SelectCode(CurDAG->getNode(SPUISD::VEC2PREFSLOT, dl, OpVT,
+ zextShuffle));
+ } else if (Opc == ISD::ADD && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
+ SDNode *CGLoad =
+ emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl));
+
+ return SelectCode(CurDAG->getNode(SPUISD::ADD64_MARKER, dl, OpVT,
+ Op.getOperand(0), Op.getOperand(1),
+ SDValue(CGLoad, 0)));
+ } else if (Opc == ISD::SUB && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
+ SDNode *CGLoad =
+ emitBuildVector(getBorrowGenerateShufMask(*CurDAG, dl));
+
+ return SelectCode(CurDAG->getNode(SPUISD::SUB64_MARKER, dl, OpVT,
+ Op.getOperand(0), Op.getOperand(1),
+ SDValue(CGLoad, 0)));
+ } else if (Opc == ISD::MUL && (OpVT == MVT::i64 || OpVT == MVT::v2i64)) {
+ SDNode *CGLoad =
+ emitBuildVector(getCarryGenerateShufMask(*CurDAG, dl));
+
+ return SelectCode(CurDAG->getNode(SPUISD::MUL64_MARKER, dl, OpVT,
+ Op.getOperand(0), Op.getOperand(1),
+ SDValue(CGLoad, 0)));
+ } else if (Opc == ISD::TRUNCATE) {
+ SDValue Op0 = Op.getOperand(0);
+ if ((Op0.getOpcode() == ISD::SRA || Op0.getOpcode() == ISD::SRL)
+ && OpVT == MVT::i32
+ && Op0.getValueType() == MVT::i64) {
+ // Catch (truncate:i32 ([sra|srl]:i64 arg, c), where c >= 32
+ //
+ // Take advantage of the fact that the upper 32 bits are in the
+ // i32 preferred slot and avoid shuffle gymnastics:
+ ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Op0.getOperand(1));
+ if (CN != 0) {
+ unsigned shift_amt = unsigned(CN->getZExtValue());
+
+ if (shift_amt >= 32) {
+ SDNode *hi32 =
+ CurDAG->getTargetNode(SPU::ORr32_r64, dl, OpVT,
+ Op0.getOperand(0));
+
+ shift_amt -= 32;
+ if (shift_amt > 0) {
+ // Take care of the additional shift, if present:
+ SDValue shift = CurDAG->getTargetConstant(shift_amt, MVT::i32);
+ unsigned Opc = SPU::ROTMAIr32_i32;
+
+ if (Op0.getOpcode() == ISD::SRL)
+ Opc = SPU::ROTMr32;
+
+ hi32 = CurDAG->getTargetNode(Opc, dl, OpVT, SDValue(hi32, 0),
+ shift);
+ }
+
+ return hi32;
+ }
}
}
+ } else if (Opc == ISD::SHL) {
+ if (OpVT == MVT::i64) {
+ return SelectSHLi64(Op, OpVT);
+ }
+ } else if (Opc == ISD::SRL) {
+ if (OpVT == MVT::i64) {
+ return SelectSRLi64(Op, OpVT);
+ }
+ } else if (Opc == ISD::SRA) {
+ if (OpVT == MVT::i64) {
+ return SelectSRAi64(Op, OpVT);
+ }
+ } else if (Opc == ISD::FNEG
+ && (OpVT == MVT::f64 || OpVT == MVT::v2f64)) {
+ DebugLoc dl = Op.getDebugLoc();
+ // Check if the pattern is a special form of DFNMS:
+ // (fneg (fsub (fmul R64FP:$rA, R64FP:$rB), R64FP:$rC))
+ SDValue Op0 = Op.getOperand(0);
+ if (Op0.getOpcode() == ISD::FSUB) {
+ SDValue Op00 = Op0.getOperand(0);
+ if (Op00.getOpcode() == ISD::FMUL) {
+ unsigned Opc = SPU::DFNMSf64;
+ if (OpVT == MVT::v2f64)
+ Opc = SPU::DFNMSv2f64;
+
+ return CurDAG->getTargetNode(Opc, dl, OpVT,
+ Op00.getOperand(0),
+ Op00.getOperand(1),
+ Op0.getOperand(1));
+ }
+ }
+
+ SDValue negConst = CurDAG->getConstant(0x8000000000000000ULL, MVT::i64);
+ SDNode *signMask = 0;
+ unsigned Opc = SPU::XORfneg64;
+
+ if (OpVT == MVT::f64) {
+ signMask = SelectI64Constant(negConst, MVT::i64, dl);
+ } else if (OpVT == MVT::v2f64) {
+ Opc = SPU::XORfnegvec;
+ signMask = emitBuildVector(CurDAG->getNode(ISD::BUILD_VECTOR, dl,
+ MVT::v2i64,
+ negConst, negConst));
+ }
+
+ return CurDAG->getTargetNode(Opc, dl, OpVT,
+ Op.getOperand(0), SDValue(signMask, 0));
+ } else if (Opc == ISD::FABS) {
+ if (OpVT == MVT::f64) {
+ SDNode *signMask = SelectI64Constant(0x7fffffffffffffffULL, MVT::i64, dl);
+ return CurDAG->getTargetNode(SPU::ANDfabs64, dl, OpVT,
+ Op.getOperand(0), SDValue(signMask, 0));
+ } else if (OpVT == MVT::v2f64) {
+ SDValue absConst = CurDAG->getConstant(0x7fffffffffffffffULL, MVT::i64);
+ SDValue absVec = CurDAG->getNode(ISD::BUILD_VECTOR, dl, MVT::v2i64,
+ absConst, absConst);
+ SDNode *signMask = emitBuildVector(absVec);
+ return CurDAG->getTargetNode(SPU::ANDfabsvec, dl, OpVT,
+ Op.getOperand(0), SDValue(signMask, 0));
+ }
} else if (Opc == SPUISD::LDRESULT) {
// Custom select instructions for LDRESULT
- MVT VT = N->getValueType(0);
+ EVT VT = N->getValueType(0);
SDValue Arg = N->getOperand(0);
SDValue Chain = N->getOperand(1);
SDNode *Result;
const valtype_map_s *vtm = getValueTypeMapEntry(VT);
if (vtm->ldresult_ins == 0) {
- cerr << "LDRESULT for unsupported type: "
- << VT.getMVTString()
- << "\n";
- abort();
+ std::string msg;
+ raw_string_ostream Msg(msg);
+ Msg << "LDRESULT for unsupported type: "
+ << VT.getEVTString();
+ llvm_report_error(Msg.str());
}
Opc = vtm->ldresult_ins;
if (vtm->ldresult_imm) {
SDValue Zero = CurDAG->getTargetConstant(0, VT);
- Result = CurDAG->getTargetNode(Opc, VT, MVT::Other, Arg, Zero, Chain);
+ Result = CurDAG->getTargetNode(Opc, dl, VT, MVT::Other, Arg, Zero, Chain);
} else {
- Result = CurDAG->getTargetNode(Opc, MVT::Other, Arg, Arg, Chain);
+ Result = CurDAG->getTargetNode(Opc, dl, VT, MVT::Other, Arg, Arg, Chain);
}
- Chain = SDValue(Result, 1);
-
return Result;
} else if (Opc == SPUISD::IndirectAddr) {
- SDValue Op0 = Op.getOperand(0);
- if (Op0.getOpcode() == SPUISD::LDRESULT) {
- /* || Op0.getOpcode() == SPUISD::AFormAddr) */
- // (IndirectAddr (LDRESULT, imm))
- SDValue Op1 = Op.getOperand(1);
- MVT VT = Op.getValueType();
-
- DEBUG(cerr << "CellSPU: IndirectAddr(LDRESULT, imm):\nOp0 = ");
- DEBUG(Op.getOperand(0).getNode()->dump(CurDAG));
- DEBUG(cerr << "\nOp1 = ");
- DEBUG(Op.getOperand(1).getNode()->dump(CurDAG));
- DEBUG(cerr << "\n");
-
+ // Look at the operands: SelectCode() will catch the cases that aren't
+ // specifically handled here.
+ //
+ // SPUInstrInfo catches the following patterns:
+ // (SPUindirect (SPUhi ...), (SPUlo ...))
+ // (SPUindirect $sp, imm)
+ EVT VT = Op.getValueType();
+ SDValue Op0 = N->getOperand(0);
+ SDValue Op1 = N->getOperand(1);
+ RegisterSDNode *RN;
+
+ if ((Op0.getOpcode() != SPUISD::Hi && Op1.getOpcode() != SPUISD::Lo)
+ || (Op0.getOpcode() == ISD::Register
+ && ((RN = dyn_cast<RegisterSDNode>(Op0.getNode())) != 0
+ && RN->getReg() != SPU::R1))) {
+ NewOpc = SPU::Ar32;
if (Op1.getOpcode() == ISD::Constant) {
ConstantSDNode *CN = cast<ConstantSDNode>(Op1);
- Op1 = CurDAG->getTargetConstant(CN->getZExtValue(), VT);
+ Op1 = CurDAG->getTargetConstant(CN->getSExtValue(), VT);
NewOpc = (isI32IntS10Immediate(CN) ? SPU::AIr32 : SPU::Ar32);
- Ops[0] = Op0;
- Ops[1] = Op1;
- n_ops = 2;
}
+ Ops[0] = Op0;
+ Ops[1] = Op1;
+ n_ops = 2;
}
}
-
+
if (n_ops > 0) {
if (N->hasOneUse())
return CurDAG->SelectNodeTo(N, NewOpc, OpVT, Ops, n_ops);
else
- return CurDAG->getTargetNode(NewOpc, OpVT, Ops, n_ops);
+ return CurDAG->getTargetNode(NewOpc, dl, OpVT, Ops, n_ops);
} else
return SelectCode(Op);
}
-/// createPPCISelDag - This pass converts a legalized DAG into a
+/*!
+ * Emit the instruction sequence for i64 left shifts. The basic algorithm
+ * is to fill the bottom two word slots with zeros so that zeros are shifted
+ * in as the entire quadword is shifted left.
+ *
+ * \note This code could also be used to implement v2i64 shl.
+ *
+ * @param Op The shl operand
+ * @param OpVT Op's machine value value type (doesn't need to be passed, but
+ * makes life easier.)
+ * @return The SDNode with the entire instruction sequence
+ */
+SDNode *
+SPUDAGToDAGISel::SelectSHLi64(SDValue &Op, EVT OpVT) {
+ SDValue Op0 = Op.getOperand(0);
+ EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
+ OpVT, (128 / OpVT.getSizeInBits()));
+ SDValue ShiftAmt = Op.getOperand(1);
+ EVT ShiftAmtVT = ShiftAmt.getValueType();
+ SDNode *VecOp0, *SelMask, *ZeroFill, *Shift = 0;
+ SDValue SelMaskVal;
+ DebugLoc dl = Op.getDebugLoc();
+
+ VecOp0 = CurDAG->getTargetNode(SPU::ORv2i64_i64, dl, VecVT, Op0);
+ SelMaskVal = CurDAG->getTargetConstant(0xff00ULL, MVT::i16);
+ SelMask = CurDAG->getTargetNode(SPU::FSMBIv2i64, dl, VecVT, SelMaskVal);
+ ZeroFill = CurDAG->getTargetNode(SPU::ILv2i64, dl, VecVT,
+ CurDAG->getTargetConstant(0, OpVT));
+ VecOp0 = CurDAG->getTargetNode(SPU::SELBv2i64, dl, VecVT,
+ SDValue(ZeroFill, 0),
+ SDValue(VecOp0, 0),
+ SDValue(SelMask, 0));
+
+ if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
+ unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
+ unsigned bits = unsigned(CN->getZExtValue()) & 7;
+
+ if (bytes > 0) {
+ Shift =
+ CurDAG->getTargetNode(SPU::SHLQBYIv2i64, dl, VecVT,
+ SDValue(VecOp0, 0),
+ CurDAG->getTargetConstant(bytes, ShiftAmtVT));
+ }
+
+ if (bits > 0) {
+ Shift =
+ CurDAG->getTargetNode(SPU::SHLQBIIv2i64, dl, VecVT,
+ SDValue((Shift != 0 ? Shift : VecOp0), 0),
+ CurDAG->getTargetConstant(bits, ShiftAmtVT));
+ }
+ } else {
+ SDNode *Bytes =
+ CurDAG->getTargetNode(SPU::ROTMIr32, dl, ShiftAmtVT,
+ ShiftAmt,
+ CurDAG->getTargetConstant(3, ShiftAmtVT));
+ SDNode *Bits =
+ CurDAG->getTargetNode(SPU::ANDIr32, dl, ShiftAmtVT,
+ ShiftAmt,
+ CurDAG->getTargetConstant(7, ShiftAmtVT));
+ Shift =
+ CurDAG->getTargetNode(SPU::SHLQBYv2i64, dl, VecVT,
+ SDValue(VecOp0, 0), SDValue(Bytes, 0));
+ Shift =
+ CurDAG->getTargetNode(SPU::SHLQBIv2i64, dl, VecVT,
+ SDValue(Shift, 0), SDValue(Bits, 0));
+ }
+
+ return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
+}
+
+/*!
+ * Emit the instruction sequence for i64 logical right shifts.
+ *
+ * @param Op The shl operand
+ * @param OpVT Op's machine value value type (doesn't need to be passed, but
+ * makes life easier.)
+ * @return The SDNode with the entire instruction sequence
+ */
+SDNode *
+SPUDAGToDAGISel::SelectSRLi64(SDValue &Op, EVT OpVT) {
+ SDValue Op0 = Op.getOperand(0);
+ EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
+ OpVT, (128 / OpVT.getSizeInBits()));
+ SDValue ShiftAmt = Op.getOperand(1);
+ EVT ShiftAmtVT = ShiftAmt.getValueType();
+ SDNode *VecOp0, *Shift = 0;
+ DebugLoc dl = Op.getDebugLoc();
+
+ VecOp0 = CurDAG->getTargetNode(SPU::ORv2i64_i64, dl, VecVT, Op0);
+
+ if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
+ unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
+ unsigned bits = unsigned(CN->getZExtValue()) & 7;
+
+ if (bytes > 0) {
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQMBYIv2i64, dl, VecVT,
+ SDValue(VecOp0, 0),
+ CurDAG->getTargetConstant(bytes, ShiftAmtVT));
+ }
+
+ if (bits > 0) {
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQMBIIv2i64, dl, VecVT,
+ SDValue((Shift != 0 ? Shift : VecOp0), 0),
+ CurDAG->getTargetConstant(bits, ShiftAmtVT));
+ }
+ } else {
+ SDNode *Bytes =
+ CurDAG->getTargetNode(SPU::ROTMIr32, dl, ShiftAmtVT,
+ ShiftAmt,
+ CurDAG->getTargetConstant(3, ShiftAmtVT));
+ SDNode *Bits =
+ CurDAG->getTargetNode(SPU::ANDIr32, dl, ShiftAmtVT,
+ ShiftAmt,
+ CurDAG->getTargetConstant(7, ShiftAmtVT));
+
+ // Ensure that the shift amounts are negated!
+ Bytes = CurDAG->getTargetNode(SPU::SFIr32, dl, ShiftAmtVT,
+ SDValue(Bytes, 0),
+ CurDAG->getTargetConstant(0, ShiftAmtVT));
+
+ Bits = CurDAG->getTargetNode(SPU::SFIr32, dl, ShiftAmtVT,
+ SDValue(Bits, 0),
+ CurDAG->getTargetConstant(0, ShiftAmtVT));
+
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQMBYv2i64, dl, VecVT,
+ SDValue(VecOp0, 0), SDValue(Bytes, 0));
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQMBIv2i64, dl, VecVT,
+ SDValue(Shift, 0), SDValue(Bits, 0));
+ }
+
+ return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
+}
+
+/*!
+ * Emit the instruction sequence for i64 arithmetic right shifts.
+ *
+ * @param Op The shl operand
+ * @param OpVT Op's machine value value type (doesn't need to be passed, but
+ * makes life easier.)
+ * @return The SDNode with the entire instruction sequence
+ */
+SDNode *
+SPUDAGToDAGISel::SelectSRAi64(SDValue &Op, EVT OpVT) {
+ // Promote Op0 to vector
+ EVT VecVT = EVT::getVectorVT(*CurDAG->getContext(),
+ OpVT, (128 / OpVT.getSizeInBits()));
+ SDValue ShiftAmt = Op.getOperand(1);
+ EVT ShiftAmtVT = ShiftAmt.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+
+ SDNode *VecOp0 =
+ CurDAG->getTargetNode(SPU::ORv2i64_i64, dl, VecVT, Op.getOperand(0));
+
+ SDValue SignRotAmt = CurDAG->getTargetConstant(31, ShiftAmtVT);
+ SDNode *SignRot =
+ CurDAG->getTargetNode(SPU::ROTMAIv2i64_i32, dl, MVT::v2i64,
+ SDValue(VecOp0, 0), SignRotAmt);
+ SDNode *UpperHalfSign =
+ CurDAG->getTargetNode(SPU::ORi32_v4i32, dl, MVT::i32, SDValue(SignRot, 0));
+
+ SDNode *UpperHalfSignMask =
+ CurDAG->getTargetNode(SPU::FSM64r32, dl, VecVT, SDValue(UpperHalfSign, 0));
+ SDNode *UpperLowerMask =
+ CurDAG->getTargetNode(SPU::FSMBIv2i64, dl, VecVT,
+ CurDAG->getTargetConstant(0xff00ULL, MVT::i16));
+ SDNode *UpperLowerSelect =
+ CurDAG->getTargetNode(SPU::SELBv2i64, dl, VecVT,
+ SDValue(UpperHalfSignMask, 0),
+ SDValue(VecOp0, 0),
+ SDValue(UpperLowerMask, 0));
+
+ SDNode *Shift = 0;
+
+ if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(ShiftAmt)) {
+ unsigned bytes = unsigned(CN->getZExtValue()) >> 3;
+ unsigned bits = unsigned(CN->getZExtValue()) & 7;
+
+ if (bytes > 0) {
+ bytes = 31 - bytes;
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQBYIv2i64, dl, VecVT,
+ SDValue(UpperLowerSelect, 0),
+ CurDAG->getTargetConstant(bytes, ShiftAmtVT));
+ }
+
+ if (bits > 0) {
+ bits = 8 - bits;
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQBIIv2i64, dl, VecVT,
+ SDValue((Shift != 0 ? Shift : UpperLowerSelect), 0),
+ CurDAG->getTargetConstant(bits, ShiftAmtVT));
+ }
+ } else {
+ SDNode *NegShift =
+ CurDAG->getTargetNode(SPU::SFIr32, dl, ShiftAmtVT,
+ ShiftAmt, CurDAG->getTargetConstant(0, ShiftAmtVT));
+
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQBYBIv2i64_r32, dl, VecVT,
+ SDValue(UpperLowerSelect, 0), SDValue(NegShift, 0));
+ Shift =
+ CurDAG->getTargetNode(SPU::ROTQBIv2i64, dl, VecVT,
+ SDValue(Shift, 0), SDValue(NegShift, 0));
+ }
+
+ return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT, SDValue(Shift, 0));
+}
+
+/*!
+ Do the necessary magic necessary to load a i64 constant
+ */
+SDNode *SPUDAGToDAGISel::SelectI64Constant(SDValue& Op, EVT OpVT,
+ DebugLoc dl) {
+ ConstantSDNode *CN = cast<ConstantSDNode>(Op.getNode());
+ return SelectI64Constant(CN->getZExtValue(), OpVT, dl);
+}
+
+SDNode *SPUDAGToDAGISel::SelectI64Constant(uint64_t Value64, EVT OpVT,
+ DebugLoc dl) {
+ EVT OpVecVT = EVT::getVectorVT(*CurDAG->getContext(), OpVT, 2);
+ SDValue i64vec =
+ SPU::LowerV2I64Splat(OpVecVT, *CurDAG, Value64, dl);
+
+ // Here's where it gets interesting, because we have to parse out the
+ // subtree handed back in i64vec:
+
+ if (i64vec.getOpcode() == ISD::BIT_CONVERT) {
+ // The degenerate case where the upper and lower bits in the splat are
+ // identical:
+ SDValue Op0 = i64vec.getOperand(0);
+
+ ReplaceUses(i64vec, Op0);
+ return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT,
+ SDValue(emitBuildVector(Op0), 0));
+ } else if (i64vec.getOpcode() == SPUISD::SHUFB) {
+ SDValue lhs = i64vec.getOperand(0);
+ SDValue rhs = i64vec.getOperand(1);
+ SDValue shufmask = i64vec.getOperand(2);
+
+ if (lhs.getOpcode() == ISD::BIT_CONVERT) {
+ ReplaceUses(lhs, lhs.getOperand(0));
+ lhs = lhs.getOperand(0);
+ }
+
+ SDNode *lhsNode = (lhs.getNode()->isMachineOpcode()
+ ? lhs.getNode()
+ : emitBuildVector(lhs));
+
+ if (rhs.getOpcode() == ISD::BIT_CONVERT) {
+ ReplaceUses(rhs, rhs.getOperand(0));
+ rhs = rhs.getOperand(0);
+ }
+
+ SDNode *rhsNode = (rhs.getNode()->isMachineOpcode()
+ ? rhs.getNode()
+ : emitBuildVector(rhs));
+
+ if (shufmask.getOpcode() == ISD::BIT_CONVERT) {
+ ReplaceUses(shufmask, shufmask.getOperand(0));
+ shufmask = shufmask.getOperand(0);
+ }
+
+ SDNode *shufMaskNode = (shufmask.getNode()->isMachineOpcode()
+ ? shufmask.getNode()
+ : emitBuildVector(shufmask));
+
+ SDNode *shufNode =
+ Select(CurDAG->getNode(SPUISD::SHUFB, dl, OpVecVT,
+ SDValue(lhsNode, 0), SDValue(rhsNode, 0),
+ SDValue(shufMaskNode, 0)));
+
+ return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT,
+ SDValue(shufNode, 0));
+ } else if (i64vec.getOpcode() == ISD::BUILD_VECTOR) {
+ return CurDAG->getTargetNode(SPU::ORi64_v2i64, dl, OpVT,
+ SDValue(emitBuildVector(i64vec), 0));
+ } else {
+ llvm_report_error("SPUDAGToDAGISel::SelectI64Constant: Unhandled i64vec"
+ "condition");
+ }
+}
+
+/// createSPUISelDag - This pass converts a legalized DAG into a
/// SPU-specific DAG, ready for instruction scheduling.
///
FunctionPass *llvm::createSPUISelDag(SPUTargetMachine &TM) {