#define DEBUG_TYPE "ppc-codegen"
#include "PPC.h"
-#include "PPCPredicates.h"
+#include "MCTargetDesc/PPCPredicates.h"
#include "PPCTargetMachine.h"
-#include "PPCISelLowering.h"
-#include "PPCHazardRecognizers.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineFunctionAnalysis.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
-#include "llvm/Target/TargetOptions.h"
-#include "llvm/Constants.h"
-#include "llvm/Function.h"
-#include "llvm/GlobalValue.h"
-#include "llvm/Intrinsics.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/GlobalAlias.h"
+#include "llvm/IR/GlobalValue.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/Intrinsics.h"
#include "llvm/Support/Debug.h"
-#include "llvm/Support/MathExtras.h"
#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetOptions.h"
using namespace llvm;
+namespace llvm {
+ void initializePPCDAGToDAGISelPass(PassRegistry&);
+}
+
namespace {
//===--------------------------------------------------------------------===//
/// PPCDAGToDAGISel - PPC specific code to select PPC machine
/// instructions for SelectionDAG operations.
///
class PPCDAGToDAGISel : public SelectionDAGISel {
- PPCTargetMachine &TM;
- PPCTargetLowering &PPCLowering;
+ const PPCTargetMachine &TM;
+ const PPCTargetLowering &PPCLowering;
const PPCSubtarget &PPCSubTarget;
unsigned GlobalBaseReg;
public:
explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
: SelectionDAGISel(tm), TM(tm),
PPCLowering(*TM.getTargetLowering()),
- PPCSubTarget(*TM.getSubtargetImpl()) {}
-
+ PPCSubTarget(*TM.getSubtargetImpl()) {
+ initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
+ }
+
virtual bool runOnMachineFunction(MachineFunction &MF) {
// Make sure we re-emit a set of the global base reg if necessary
GlobalBaseReg = 0;
SelectionDAGISel::runOnMachineFunction(MF);
-
- InsertVRSaveCode(MF);
+
+ if (!PPCSubTarget.isSVR4ABI())
+ InsertVRSaveCode(MF);
+
return true;
}
-
+
+ virtual void PostprocessISelDAG();
+
/// getI32Imm - Return a target constant with the specified value, of type
/// i32.
inline SDValue getI32Imm(unsigned Imm) {
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, PPCLowering.getPointerTy());
}
-
- /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
+
+ /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
/// with any number of 0s on either side. The 1s are allowed to wrap from
/// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.
/// 0x0F0F0000 is not, since all 1s are not contiguous.
/// rotate and mask opcode and mask operation.
static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
unsigned &SH, unsigned &MB, unsigned &ME);
-
+
/// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC
/// base register. Return the virtual register that holds this value.
SDNode *getGlobalBaseReg();
-
+
// Select - Convert the specified operand from a target-independent to a
// target-specific node if it hasn't already been changed.
SDNode *Select(SDNode *N);
-
+
SDNode *SelectBitfieldInsert(SDNode *N);
/// SelectCC - Select a comparison of the specified values with the
/// SelectAddrImm - Returns true if the address N can be represented by
/// a base register plus a signed 16-bit displacement [r+imm].
- bool SelectAddrImm(SDNode *Op, SDValue N, SDValue &Disp,
+ bool SelectAddrImm(SDValue N, SDValue &Disp,
SDValue &Base) {
return PPCLowering.SelectAddressRegImm(N, Disp, Base, *CurDAG);
}
-
+
/// SelectAddrImmOffs - Return true if the operand is valid for a preinc
- /// immediate field. Because preinc imms have already been validated, just
- /// accept it.
- bool SelectAddrImmOffs(SDNode *Op, SDValue N, SDValue &Out) const {
- Out = N;
- return true;
+ /// immediate field. Note that the operand at this point is already the
+ /// result of a prior SelectAddressRegImm call.
+ bool SelectAddrImmOffs(SDValue N, SDValue &Out) const {
+ if (N.getOpcode() == ISD::TargetConstant ||
+ N.getOpcode() == ISD::TargetGlobalAddress) {
+ Out = N;
+ return true;
+ }
+
+ return false;
}
-
+
/// SelectAddrIdx - Given the specified addressed, check to see if it can be
/// represented as an indexed [r+r] operation. Returns false if it can
/// be represented by [r+imm], which are preferred.
- bool SelectAddrIdx(SDNode *Op, SDValue N, SDValue &Base,
- SDValue &Index) {
+ bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) {
return PPCLowering.SelectAddressRegReg(N, Base, Index, *CurDAG);
}
-
+
/// SelectAddrIdxOnly - Given the specified addressed, force it to be
/// represented as an indexed [r+r] operation.
- bool SelectAddrIdxOnly(SDNode *Op, SDValue N, SDValue &Base,
- SDValue &Index) {
+ bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) {
return PPCLowering.SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
}
/// SelectAddrImmShift - Returns true if the address N can be represented by
/// a base register plus a signed 14-bit displacement [r+imm*4]. Suitable
/// for use by STD and friends.
- bool SelectAddrImmShift(SDNode *Op, SDValue N, SDValue &Disp,
- SDValue &Base) {
+ bool SelectAddrImmShift(SDValue N, SDValue &Disp, SDValue &Base) {
return PPCLowering.SelectAddressRegImmShift(N, Disp, Base, *CurDAG);
}
-
+
+ // Select an address into a single register.
+ bool SelectAddr(SDValue N, SDValue &Base) {
+ Base = N;
+ return true;
+ }
+
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
/// inline asm expressions. It is always correct to compute the value into
/// a register. The case of adding a (possibly relocatable) constant to a
OutOps.push_back(Op);
return false;
}
-
- SDValue BuildSDIVSequence(SDNode *N);
- SDValue BuildUDIVSequence(SDNode *N);
-
+
void InsertVRSaveCode(MachineFunction &MF);
virtual const char *getPassName() const {
return "PowerPC DAG->DAG Pattern Instruction Selection";
- }
-
- /// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
- /// this target when scheduling the DAG.
- virtual ScheduleHazardRecognizer *CreateTargetHazardRecognizer() {
- // Should use subtarget info to pick the right hazard recognizer. For
- // now, always return a PPC970 recognizer.
- const TargetInstrInfo *II = TM.getInstrInfo();
- assert(II && "No InstrInfo?");
- return new PPCHazardRecognizer970(*II);
}
// Include the pieces autogenerated from the target description.
#include "PPCGenDAGISel.inc"
-
+
private:
SDNode *SelectSETCC(SDNode *N);
};
/// check to see if we need to save/restore VRSAVE. If so, do it.
void PPCDAGToDAGISel::InsertVRSaveCode(MachineFunction &Fn) {
// Check to see if this function uses vector registers, which means we have to
- // save and restore the VRSAVE register and update it with the regs we use.
+ // save and restore the VRSAVE register and update it with the regs we use.
//
// In this case, there will be virtual registers of vector type created
// by the scheduler. Detect them now.
bool HasVectorVReg = false;
- for (unsigned i = TargetRegisterInfo::FirstVirtualRegister,
- e = RegInfo->getLastVirtReg()+1; i != e; ++i)
- if (RegInfo->getRegClass(i) == &PPC::VRRCRegClass) {
+ for (unsigned i = 0, e = RegInfo->getNumVirtRegs(); i != e; ++i) {
+ unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
+ if (RegInfo->getRegClass(Reg) == &PPC::VRRCRegClass) {
HasVectorVReg = true;
break;
}
+ }
if (!HasVectorVReg) return; // nothing to do.
-
+
// If we have a vector register, we want to emit code into the entry and exit
// blocks to save and restore the VRSAVE register. We do this here (instead
// of marking all vector instructions as clobbering VRSAVE) for two reasons:
// function and one for the value after having bits or'd into it.
unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
-
+
const TargetInstrInfo &TII = *TM.getInstrInfo();
MachineBasicBlock &EntryBB = *Fn.begin();
- DebugLoc dl = DebugLoc::getUnknownLoc();
+ DebugLoc dl;
// Emit the following code into the entry block:
// InVRSAVE = MFVRSAVE
// UpdatedVRSAVE = UPDATE_VRSAVE InVRSAVE
BuildMI(EntryBB, IP, dl, TII.get(PPC::UPDATE_VRSAVE),
UpdatedVRSAVE).addReg(InVRSAVE);
BuildMI(EntryBB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(UpdatedVRSAVE);
-
+
// Find all return blocks, outputting a restore in each epilog.
for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
- if (!BB->empty() && BB->back().getDesc().isReturn()) {
+ if (!BB->empty() && BB->back().isReturn()) {
IP = BB->end(); --IP;
-
+
// Skip over all terminator instructions, which are part of the return
// sequence.
MachineBasicBlock::iterator I2 = IP;
- while (I2 != BB->begin() && (--I2)->getDesc().isTerminator())
+ while (I2 != BB->begin() && (--I2)->isTerminator())
IP = I2;
-
+
// Emit: MTVRSAVE InVRSave
BuildMI(*BB, IP, dl, TII.get(PPC::MTVRSAVE)).addReg(InVRSAVE);
- }
+ }
}
}
// Insert the set of GlobalBaseReg into the first MBB of the function
MachineBasicBlock &FirstMBB = MF->front();
MachineBasicBlock::iterator MBBI = FirstMBB.begin();
- DebugLoc dl = DebugLoc::getUnknownLoc();
+ DebugLoc dl;
if (PPCLowering.getPointerTy() == MVT::i32) {
- GlobalBaseReg = RegInfo->createVirtualRegister(PPC::GPRCRegisterClass);
- BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR), PPC::LR);
+ GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
} else {
- GlobalBaseReg = RegInfo->createVirtualRegister(PPC::G8RCRegisterClass);
- BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8), PPC::LR8);
+ GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RCRegClass);
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
}
}
return false;
}
-bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
- bool isShiftMask, unsigned &SH,
+bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
+ bool isShiftMask, unsigned &SH,
unsigned &MB, unsigned &ME) {
// Don't even go down this path for i64, since different logic will be
// necessary for rldicl/rldicr/rldimi.
if (N->getNumOperands() != 2 ||
!isInt32Immediate(N->getOperand(1).getNode(), Shift) || (Shift > 31))
return false;
-
+
if (Opcode == ISD::SHL) {
// apply shift left to mask if it comes first
if (isShiftMask) Mask = Mask << Shift;
// determine which bits are made indeterminant by shift
Indeterminant = ~(0xFFFFFFFFu << Shift);
- } else if (Opcode == ISD::SRL) {
+ } else if (Opcode == ISD::SRL) {
// apply shift right to mask if it comes first
if (isShiftMask) Mask = Mask >> Shift;
// determine which bits are made indeterminant by shift
} else {
return false;
}
-
+
// if the mask doesn't intersect any Indeterminant bits
if (Mask && !(Mask & Indeterminant)) {
SH = Shift & 31;
SDValue Op0 = N->getOperand(0);
SDValue Op1 = N->getOperand(1);
DebugLoc dl = N->getDebugLoc();
-
+
APInt LKZ, LKO, RKZ, RKO;
- CurDAG->ComputeMaskedBits(Op0, APInt::getAllOnesValue(32), LKZ, LKO);
- CurDAG->ComputeMaskedBits(Op1, APInt::getAllOnesValue(32), RKZ, RKO);
-
+ CurDAG->ComputeMaskedBits(Op0, LKZ, LKO);
+ CurDAG->ComputeMaskedBits(Op1, RKZ, RKO);
+
unsigned TargetMask = LKZ.getZExtValue();
unsigned InsertMask = RKZ.getZExtValue();
-
+
if ((TargetMask | InsertMask) == 0xFFFFFFFF) {
unsigned Op0Opc = Op0.getOpcode();
unsigned Op1Opc = Op1.getOpcode();
std::swap(TargetMask, InsertMask);
}
}
-
+
unsigned MB, ME;
if (InsertMask && isRunOfOnes(InsertMask, MB, ME)) {
SDValue Tmp1, Tmp2;
ISD::CondCode CC, DebugLoc dl) {
// Always select the LHS.
unsigned Opc;
-
+
if (LHS.getValueType() == MVT::i32) {
unsigned Imm;
if (CC == ISD::SETEQ || CC == ISD::SETNE) {
if (isInt32Immediate(RHS, Imm)) {
// SETEQ/SETNE comparison with 16-bit immediate, fold it.
- if (isUInt16(Imm))
+ if (isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
getI32Imm(Imm & 0xFFFF)), 0);
// If this is a 16-bit signed immediate, fold it.
- if (isInt16((int)Imm))
+ if (isInt<16>((int)Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
getI32Imm(Imm & 0xFFFF)), 0);
-
+
// For non-equality comparisons, the default code would materialize the
// constant, then compare against it, like this:
// lis r2, 4660
- // ori r2, r2, 22136
+ // ori r2, r2, 22136
// cmpw cr0, r3, r2
// Since we are just comparing for equality, we can emit this instead:
// xoris r0,r3,0x1234
}
Opc = PPC::CMPLW;
} else if (ISD::isUnsignedIntSetCC(CC)) {
- if (isInt32Immediate(RHS, Imm) && isUInt16(Imm))
+ if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
getI32Imm(Imm & 0xFFFF)), 0);
Opc = PPC::CMPLW;
if (CC == ISD::SETEQ || CC == ISD::SETNE) {
if (isInt64Immediate(RHS.getNode(), Imm)) {
// SETEQ/SETNE comparison with 16-bit immediate, fold it.
- if (isUInt16(Imm))
+ if (isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
getI32Imm(Imm & 0xFFFF)), 0);
// If this is a 16-bit signed immediate, fold it.
- if (isInt16(Imm))
+ if (isInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
getI32Imm(Imm & 0xFFFF)), 0);
-
+
// For non-equality comparisons, the default code would materialize the
// constant, then compare against it, like this:
// lis r2, 4660
- // ori r2, r2, 22136
+ // ori r2, r2, 22136
// cmpd cr0, r3, r2
// Since we are just comparing for equality, we can emit this instead:
// xoris r0,r3,0x1234
// cmpldi cr0,r0,0x5678
// beq cr0,L6
- if (isUInt32(Imm)) {
+ if (isUInt<32>(Imm)) {
SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
getI64Imm(Imm >> 16)), 0);
return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
}
Opc = PPC::CMPLD;
} else if (ISD::isUnsignedIntSetCC(CC)) {
- if (isInt64Immediate(RHS.getNode(), Imm) && isUInt16(Imm))
+ if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
getI64Imm(Imm & 0xFFFF)), 0);
Opc = PPC::CMPLD;
case ISD::SETUNE:
case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE
case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO
- case ISD::SETUEQ:
- case ISD::SETOGE:
- case ISD::SETOLE:
+ case ISD::SETUEQ:
+ case ISD::SETOGE:
+ case ISD::SETOLE:
case ISD::SETONE:
llvm_unreachable("Invalid branch code: should be expanded by legalize");
// These are invalid for floating point. Assume integer.
case ISD::SETULT: return 0;
case ISD::SETUGT: return 1;
}
- return 0;
}
+// getVCmpInst: return the vector compare instruction for the specified
+// vector type and condition code. Since this is for altivec specific code,
+// only support the altivec types (v16i8, v8i16, v4i32, and v4f32).
+static unsigned int getVCmpInst(MVT::SimpleValueType VecVT, ISD::CondCode CC) {
+ switch (CC) {
+ case ISD::SETEQ:
+ case ISD::SETUEQ:
+ case ISD::SETNE:
+ case ISD::SETUNE:
+ if (VecVT == MVT::v16i8)
+ return PPC::VCMPEQUB;
+ else if (VecVT == MVT::v8i16)
+ return PPC::VCMPEQUH;
+ else if (VecVT == MVT::v4i32)
+ return PPC::VCMPEQUW;
+ // v4f32 != v4f32 could be translate to unordered not equal
+ else if (VecVT == MVT::v4f32)
+ return PPC::VCMPEQFP;
+ break;
+ case ISD::SETLT:
+ case ISD::SETGT:
+ case ISD::SETLE:
+ case ISD::SETGE:
+ if (VecVT == MVT::v16i8)
+ return PPC::VCMPGTSB;
+ else if (VecVT == MVT::v8i16)
+ return PPC::VCMPGTSH;
+ else if (VecVT == MVT::v4i32)
+ return PPC::VCMPGTSW;
+ else if (VecVT == MVT::v4f32)
+ return PPC::VCMPGTFP;
+ break;
+ case ISD::SETULT:
+ case ISD::SETUGT:
+ case ISD::SETUGE:
+ case ISD::SETULE:
+ if (VecVT == MVT::v16i8)
+ return PPC::VCMPGTUB;
+ else if (VecVT == MVT::v8i16)
+ return PPC::VCMPGTUH;
+ else if (VecVT == MVT::v4i32)
+ return PPC::VCMPGTUW;
+ break;
+ case ISD::SETOEQ:
+ if (VecVT == MVT::v4f32)
+ return PPC::VCMPEQFP;
+ break;
+ case ISD::SETOLT:
+ case ISD::SETOGT:
+ case ISD::SETOLE:
+ if (VecVT == MVT::v4f32)
+ return PPC::VCMPGTFP;
+ break;
+ case ISD::SETOGE:
+ if (VecVT == MVT::v4f32)
+ return PPC::VCMPGEFP;
+ break;
+ default:
+ break;
+ }
+ llvm_unreachable("Invalid integer vector compare condition");
+}
+
+// getVCmpEQInst: return the equal compare instruction for the specified vector
+// type. Since this is for altivec specific code, only support the altivec
+// types (v16i8, v8i16, v4i32, and v4f32).
+static unsigned int getVCmpEQInst(MVT::SimpleValueType VecVT) {
+ switch (VecVT) {
+ case MVT::v16i8:
+ return PPC::VCMPEQUB;
+ case MVT::v8i16:
+ return PPC::VCMPEQUH;
+ case MVT::v4i32:
+ return PPC::VCMPEQUW;
+ case MVT::v4f32:
+ return PPC::VCMPEQFP;
+ default:
+ llvm_unreachable("Invalid integer vector compare condition");
+ }
+}
+
+
SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) {
DebugLoc dl = N->getDebugLoc();
unsigned Imm;
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
+ EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
+ bool isPPC64 = (PtrVT == MVT::i64);
+
if (isInt32Immediate(N->getOperand(1), Imm)) {
// We can codegen setcc op, imm very efficiently compared to a brcond.
// Check for those cases here.
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
}
case ISD::SETNE: {
+ if (isPPC64) break;
SDValue AD =
- SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Flag,
+ SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
Op, getI32Imm(~0U)), 0);
- return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
+ return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
AD.getValue(1));
}
case ISD::SETLT: {
switch (CC) {
default: break;
case ISD::SETEQ:
- Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Flag,
+ if (isPPC64) break;
+ Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
Op, getI32Imm(1)), 0);
- return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
- SDValue(CurDAG->getMachineNode(PPC::LI, dl,
+ return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
+ SDValue(CurDAG->getMachineNode(PPC::LI, dl,
MVT::i32,
getI32Imm(0)), 0),
Op.getValue(1));
case ISD::SETNE: {
+ if (isPPC64) break;
Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
- SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Flag,
+ SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
Op, getI32Imm(~0U));
return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0),
Op, SDValue(AD, 1));
}
case ISD::SETGT: {
SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
- Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4),
+ Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4),
0);
- return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
+ return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
getI32Imm(1));
}
}
}
}
-
+
+ SDValue LHS = N->getOperand(0);
+ SDValue RHS = N->getOperand(1);
+
+ // Altivec Vector compare instructions do not set any CR register by default and
+ // vector compare operations return the same type as the operands.
+ if (LHS.getValueType().isVector()) {
+ EVT VecVT = LHS.getValueType();
+ MVT::SimpleValueType VT = VecVT.getSimpleVT().SimpleTy;
+ unsigned int VCmpInst = getVCmpInst(VT, CC);
+
+ switch (CC) {
+ case ISD::SETEQ:
+ case ISD::SETOEQ:
+ case ISD::SETUEQ:
+ return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
+ case ISD::SETNE:
+ case ISD::SETONE:
+ case ISD::SETUNE: {
+ SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
+ return CurDAG->SelectNodeTo(N, PPC::VNOR, VecVT, VCmp, VCmp);
+ }
+ case ISD::SETLT:
+ case ISD::SETOLT:
+ case ISD::SETULT:
+ return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, RHS, LHS);
+ case ISD::SETGT:
+ case ISD::SETOGT:
+ case ISD::SETUGT:
+ return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
+ case ISD::SETGE:
+ case ISD::SETOGE:
+ case ISD::SETUGE: {
+ // Small optimization: Altivec provides a 'Vector Compare Greater Than
+ // or Equal To' instruction (vcmpgefp), so in this case there is no
+ // need for extra logic for the equal compare.
+ if (VecVT.getSimpleVT().isFloatingPoint()) {
+ return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
+ } else {
+ SDValue VCmpGT(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
+ unsigned int VCmpEQInst = getVCmpEQInst(VT);
+ SDValue VCmpEQ(CurDAG->getMachineNode(VCmpEQInst, dl, VecVT, LHS, RHS), 0);
+ return CurDAG->SelectNodeTo(N, PPC::VOR, VecVT, VCmpGT, VCmpEQ);
+ }
+ }
+ case ISD::SETLE:
+ case ISD::SETOLE:
+ case ISD::SETULE: {
+ SDValue VCmpLE(CurDAG->getMachineNode(VCmpInst, dl, VecVT, RHS, LHS), 0);
+ unsigned int VCmpEQInst = getVCmpEQInst(VT);
+ SDValue VCmpEQ(CurDAG->getMachineNode(VCmpEQInst, dl, VecVT, LHS, RHS), 0);
+ return CurDAG->SelectNodeTo(N, PPC::VOR, VecVT, VCmpLE, VCmpEQ);
+ }
+ default:
+ llvm_unreachable("Invalid vector compare type: should be expanded by legalize");
+ }
+ }
+
bool Inv;
int OtherCondIdx;
unsigned Idx = getCRIdxForSetCC(CC, Inv, OtherCondIdx);
- SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
+ SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
SDValue IntCR;
-
+
// Force the ccreg into CR7.
SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
-
+
SDValue InFlag(0, 0); // Null incoming flag value.
- CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
+ CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
InFlag).getValue(1);
-
- if (PPCSubTarget.isGigaProcessor() && OtherCondIdx == -1)
+
+ if (PPCSubTarget.hasMFOCRF() && OtherCondIdx == -1)
IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
CCReg), 0);
else
- IntCR = SDValue(CurDAG->getMachineNode(PPC::MFCR, dl, MVT::i32, CCReg), 0);
-
+ IntCR = SDValue(CurDAG->getMachineNode(PPC::MFCRpseud, dl, MVT::i32,
+ CR7Reg, CCReg), 0);
+
SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
getI32Imm(31), getI32Imm(31) };
if (OtherCondIdx == -1 && !Inv)
// Get the other bit of the comparison.
Ops[1] = getI32Imm((32-(3-OtherCondIdx)) & 31);
- SDValue OtherCond =
+ SDValue OtherCond =
SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4), 0);
return CurDAG->SelectNodeTo(N, PPC::OR, MVT::i32, Tmp, OtherCond);
switch (N->getOpcode()) {
default: break;
-
+
case ISD::Constant: {
if (N->getValueType(0) == MVT::i64) {
// Get 64 bit value.
unsigned Remainder = 0;
// Assume no shift required.
unsigned Shift = 0;
-
+
// If it can't be represented as a 32 bit value.
- if (!isInt32(Imm)) {
+ if (!isInt<32>(Imm)) {
Shift = CountTrailingZeros_64(Imm);
int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
-
+
// If the shifted value fits 32 bits.
- if (isInt32(ImmSh)) {
+ if (isInt<32>(ImmSh)) {
// Go with the shifted value.
Imm = ImmSh;
} else {
Imm >>= 32;
}
}
-
+
// Intermediate operand.
SDNode *Result;
// Handle first 32 bits.
unsigned Lo = Imm & 0xFFFF;
unsigned Hi = (Imm >> 16) & 0xFFFF;
-
+
// Simple value.
- if (isInt16(Imm)) {
+ if (isInt<16>(Imm)) {
// Just the Lo bits.
Result = CurDAG->getMachineNode(PPC::LI8, dl, MVT::i64, getI32Imm(Lo));
} else if (Lo) {
// Just the Hi bits.
Result = CurDAG->getMachineNode(PPC::LIS8, dl, MVT::i64, getI32Imm(Hi));
}
-
+
// If no shift, we're done.
if (!Shift) return Result;
if ((Hi = (Remainder >> 16) & 0xFFFF)) {
Result = CurDAG->getMachineNode(PPC::ORIS8, dl, MVT::i64,
SDValue(Result, 0), getI32Imm(Hi));
- }
+ }
if ((Lo = Remainder & 0xFFFF)) {
Result = CurDAG->getMachineNode(PPC::ORI8, dl, MVT::i64,
SDValue(Result, 0), getI32Imm(Lo));
}
-
+
return Result;
}
break;
}
-
+
case ISD::SETCC:
return SelectSETCC(N);
case PPCISD::GlobalBaseReg:
return getGlobalBaseReg();
-
+
case ISD::FrameIndex: {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
case PPCISD::MFCR: {
SDValue InFlag = N->getOperand(1);
// Use MFOCRF if supported.
- if (PPCSubTarget.isGigaProcessor())
+ if (PPCSubTarget.hasMFOCRF())
return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
N->getOperand(0), InFlag);
else
- return CurDAG->getMachineNode(PPC::MFCR, dl, MVT::i32, InFlag);
+ return CurDAG->getMachineNode(PPC::MFCRpseud, dl, MVT::i32,
+ N->getOperand(0), InFlag);
}
-
+
case ISD::SDIV: {
// FIXME: since this depends on the setting of the carry flag from the srawi
// we should really be making notes about that for the scheduler.
- // FIXME: It sure would be nice if we could cheaply recognize the
+ // FIXME: It sure would be nice if we could cheaply recognize the
// srl/add/sra pattern the dag combiner will generate for this as
// sra/addze rather than having to handle sdiv ourselves. oh well.
unsigned Imm;
SDValue N0 = N->getOperand(0);
if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
SDNode *Op =
- CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Flag,
+ CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
N0, getI32Imm(Log2_32(Imm)));
- return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
+ return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
SDValue(Op, 0), SDValue(Op, 1));
} else if ((signed)Imm < 0 && isPowerOf2_32(-Imm)) {
SDNode *Op =
- CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Flag,
+ CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
N0, getI32Imm(Log2_32(-Imm)));
SDValue PT =
SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32,
return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
}
}
-
+
// Other cases are autogenerated.
break;
}
-
+
case ISD::LOAD: {
// Handle preincrement loads.
LoadSDNode *LD = cast<LoadSDNode>(N);
EVT LoadedVT = LD->getMemoryVT();
-
+
// Normal loads are handled by code generated from the .td file.
if (LD->getAddressingMode() != ISD::PRE_INC)
break;
-
+
SDValue Offset = LD->getOffset();
- if (isa<ConstantSDNode>(Offset) ||
+ if (Offset.getOpcode() == ISD::TargetConstant ||
Offset.getOpcode() == ISD::TargetGlobalAddress) {
-
+
unsigned Opcode;
bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
if (LD->getValueType(0) != MVT::i64) {
case MVT::i8: Opcode = PPC::LBZU8; break;
}
}
-
+
SDValue Chain = LD->getChain();
SDValue Base = LD->getBasePtr();
SDValue Ops[] = { Offset, Base, Chain };
- // FIXME: PPC64
return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
PPCLowering.getPointerTy(),
MVT::Other, Ops, 3);
} else {
- llvm_unreachable("R+R preindex loads not supported yet!");
+ unsigned Opcode;
+ bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
+ if (LD->getValueType(0) != MVT::i64) {
+ // Handle PPC32 integer and normal FP loads.
+ assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
+ switch (LoadedVT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("Invalid PPC load type!");
+ case MVT::f64: Opcode = PPC::LFDUX; break;
+ case MVT::f32: Opcode = PPC::LFSUX; break;
+ case MVT::i32: Opcode = PPC::LWZUX; break;
+ case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break;
+ case MVT::i1:
+ case MVT::i8: Opcode = PPC::LBZUX; break;
+ }
+ } else {
+ assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!");
+ assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) &&
+ "Invalid sext update load");
+ switch (LoadedVT.getSimpleVT().SimpleTy) {
+ default: llvm_unreachable("Invalid PPC load type!");
+ case MVT::i64: Opcode = PPC::LDUX; break;
+ case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break;
+ case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break;
+ case MVT::i1:
+ case MVT::i8: Opcode = PPC::LBZUX8; break;
+ }
+ }
+
+ SDValue Chain = LD->getChain();
+ SDValue Base = LD->getBasePtr();
+ SDValue Ops[] = { Base, Offset, Chain };
+ return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
+ PPCLowering.getPointerTy(),
+ MVT::Other, Ops, 3);
}
}
-
+
case ISD::AND: {
unsigned Imm, Imm2, SH, MB, ME;
+ uint64_t Imm64;
// If this is an and of a value rotated between 0 and 31 bits and then and'd
// with a mask, emit rlwinm
// If this is just a masked value where the input is not handled above, and
// is not a rotate-left (handled by a pattern in the .td file), emit rlwinm
if (isInt32Immediate(N->getOperand(1), Imm) &&
- isRunOfOnes(Imm, MB, ME) &&
+ isRunOfOnes(Imm, MB, ME) &&
N->getOperand(0).getOpcode() != ISD::ROTL) {
SDValue Val = N->getOperand(0);
SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
}
+ // If this is a 64-bit zero-extension mask, emit rldicl.
+ if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
+ isMask_64(Imm64)) {
+ SDValue Val = N->getOperand(0);
+ MB = 64 - CountTrailingOnes_64(Imm64);
+ SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB) };
+ return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops, 3);
+ }
// AND X, 0 -> 0, not "rlwinm 32".
if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
ReplaceUses(SDValue(N, 0), N->getOperand(1));
}
// ISD::OR doesn't get all the bitfield insertion fun.
// (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
- if (isInt32Immediate(N->getOperand(1), Imm) &&
+ if (isInt32Immediate(N->getOperand(1), Imm) &&
N->getOperand(0).getOpcode() == ISD::OR &&
isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
unsigned MB, ME;
return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops, 5);
}
}
-
+
// Other cases are autogenerated.
break;
}
if (N->getValueType(0) == MVT::i32)
if (SDNode *I = SelectBitfieldInsert(N))
return I;
-
+
// Other cases are autogenerated.
break;
case ISD::SHL: {
getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
}
-
+
// Other cases are autogenerated.
break;
}
case ISD::SRL: {
unsigned Imm, SH, MB, ME;
if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
- isRotateAndMask(N, Imm, true, SH, MB, ME)) {
+ isRotateAndMask(N, Imm, true, SH, MB, ME)) {
SDValue Ops[] = { N->getOperand(0).getOperand(0),
getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
}
-
+
// Other cases are autogenerated.
break;
}
case ISD::SELECT_CC: {
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
-
+ EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
+ bool isPPC64 = (PtrVT == MVT::i64);
+
// Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
- if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
- if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
- if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
- if (N1C->isNullValue() && N3C->isNullValue() &&
- N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
- // FIXME: Implement this optzn for PPC64.
- N->getValueType(0) == MVT::i32) {
- SDNode *Tmp =
- CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Flag,
- N->getOperand(0), getI32Imm(~0U));
- return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
- SDValue(Tmp, 0), N->getOperand(0),
- SDValue(Tmp, 1));
- }
+ if (!isPPC64)
+ if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
+ if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N->getOperand(2)))
+ if (ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N->getOperand(3)))
+ if (N1C->isNullValue() && N3C->isNullValue() &&
+ N2C->getZExtValue() == 1ULL && CC == ISD::SETNE &&
+ // FIXME: Implement this optzn for PPC64.
+ N->getValueType(0) == MVT::i32) {
+ SDNode *Tmp =
+ CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
+ N->getOperand(0), getI32Imm(~0U));
+ return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
+ SDValue(Tmp, 0), N->getOperand(0),
+ SDValue(Tmp, 1));
+ }
SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
unsigned BROpc = getPredicateForSetCC(CC);
case ISD::BR_CC: {
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
- SDValue Ops[] = { getI32Imm(getPredicateForSetCC(CC)), CondCode,
+ SDValue Ops[] = { getI32Imm(getPredicateForSetCC(CC)), CondCode,
N->getOperand(4), N->getOperand(0) };
return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops, 4);
}
SDValue Chain = N->getOperand(0);
SDValue Target = N->getOperand(1);
unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8;
- Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Target,
+ unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8;
+ Chain = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, Target,
Chain), 0);
- return CurDAG->SelectNodeTo(N, PPC::BCTR, MVT::Other, Chain);
+ return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
+ }
+ case PPCISD::TOC_ENTRY: {
+ assert (PPCSubTarget.isPPC64() && "Only supported for 64-bit ABI");
+
+ // For medium and large code model, we generate two instructions as
+ // described below. Otherwise we allow SelectCodeCommon to handle this,
+ // selecting one of LDtoc, LDtocJTI, and LDtocCPT.
+ CodeModel::Model CModel = TM.getCodeModel();
+ if (CModel != CodeModel::Medium && CModel != CodeModel::Large)
+ break;
+
+ // The first source operand is a TargetGlobalAddress or a
+ // TargetJumpTable. If it is an externally defined symbol, a symbol
+ // with common linkage, a function address, or a jump table address,
+ // or if we are generating code for large code model, we generate:
+ // LDtocL(<ga:@sym>, ADDIStocHA(%X2, <ga:@sym>))
+ // Otherwise we generate:
+ // ADDItocL(ADDIStocHA(%X2, <ga:@sym>), <ga:@sym>)
+ SDValue GA = N->getOperand(0);
+ SDValue TOCbase = N->getOperand(1);
+ SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
+ TOCbase, GA);
+
+ if (isa<JumpTableSDNode>(GA) || CModel == CodeModel::Large)
+ return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
+ SDValue(Tmp, 0));
+
+ if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
+ const GlobalValue *GValue = G->getGlobal();
+ const GlobalAlias *GAlias = dyn_cast<GlobalAlias>(GValue);
+ const GlobalValue *RealGValue = GAlias ?
+ GAlias->resolveAliasedGlobal(false) : GValue;
+ const GlobalVariable *GVar = dyn_cast<GlobalVariable>(RealGValue);
+ assert((GVar || isa<Function>(RealGValue)) &&
+ "Unexpected global value subclass!");
+
+ // An external variable is one without an initializer. For these,
+ // for variables with common linkage, and for Functions, generate
+ // the LDtocL form.
+ if (!GVar || !GVar->hasInitializer() || RealGValue->hasCommonLinkage() ||
+ RealGValue->hasAvailableExternallyLinkage())
+ return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
+ SDValue(Tmp, 0));
+ }
+
+ return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
+ SDValue(Tmp, 0), GA);
+ }
+ case PPCISD::VADD_SPLAT: {
+ // This expands into one of three sequences, depending on whether
+ // the first operand is odd or even, positive or negative.
+ assert(isa<ConstantSDNode>(N->getOperand(0)) &&
+ isa<ConstantSDNode>(N->getOperand(1)) &&
+ "Invalid operand on VADD_SPLAT!");
+
+ int Elt = N->getConstantOperandVal(0);
+ int EltSize = N->getConstantOperandVal(1);
+ unsigned Opc1, Opc2, Opc3;
+ EVT VT;
+
+ if (EltSize == 1) {
+ Opc1 = PPC::VSPLTISB;
+ Opc2 = PPC::VADDUBM;
+ Opc3 = PPC::VSUBUBM;
+ VT = MVT::v16i8;
+ } else if (EltSize == 2) {
+ Opc1 = PPC::VSPLTISH;
+ Opc2 = PPC::VADDUHM;
+ Opc3 = PPC::VSUBUHM;
+ VT = MVT::v8i16;
+ } else {
+ assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!");
+ Opc1 = PPC::VSPLTISW;
+ Opc2 = PPC::VADDUWM;
+ Opc3 = PPC::VSUBUWM;
+ VT = MVT::v4i32;
+ }
+
+ if ((Elt & 1) == 0) {
+ // Elt is even, in the range [-32,-18] + [16,30].
+ //
+ // Convert: VADD_SPLAT elt, size
+ // Into: tmp = VSPLTIS[BHW] elt
+ // VADDU[BHW]M tmp, tmp
+ // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4
+ SDValue EltVal = getI32Imm(Elt >> 1);
+ SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
+ SDValue TmpVal = SDValue(Tmp, 0);
+ return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
+
+ } else if (Elt > 0) {
+ // Elt is odd and positive, in the range [17,31].
+ //
+ // Convert: VADD_SPLAT elt, size
+ // Into: tmp1 = VSPLTIS[BHW] elt-16
+ // tmp2 = VSPLTIS[BHW] -16
+ // VSUBU[BHW]M tmp1, tmp2
+ SDValue EltVal = getI32Imm(Elt - 16);
+ SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
+ EltVal = getI32Imm(-16);
+ SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
+ return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
+ SDValue(Tmp2, 0));
+
+ } else {
+ // Elt is odd and negative, in the range [-31,-17].
+ //
+ // Convert: VADD_SPLAT elt, size
+ // Into: tmp1 = VSPLTIS[BHW] elt+16
+ // tmp2 = VSPLTIS[BHW] -16
+ // VADDU[BHW]M tmp1, tmp2
+ SDValue EltVal = getI32Imm(Elt + 16);
+ SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
+ EltVal = getI32Imm(-16);
+ SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
+ return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
+ SDValue(Tmp2, 0));
+ }
}
}
-
+
return SelectCode(N);
}
+/// PostProcessISelDAG - Perform some late peephole optimizations
+/// on the DAG representation.
+void PPCDAGToDAGISel::PostprocessISelDAG() {
+
+ // Skip peepholes at -O0.
+ if (TM.getOptLevel() == CodeGenOpt::None)
+ return;
+
+ // These optimizations are currently supported only for 64-bit SVR4.
+ if (PPCSubTarget.isDarwin() || !PPCSubTarget.isPPC64())
+ return;
+
+ SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
+ ++Position;
+
+ while (Position != CurDAG->allnodes_begin()) {
+ SDNode *N = --Position;
+ // Skip dead nodes and any non-machine opcodes.
+ if (N->use_empty() || !N->isMachineOpcode())
+ continue;
+
+ unsigned FirstOp;
+ unsigned StorageOpcode = N->getMachineOpcode();
+
+ switch (StorageOpcode) {
+ default: continue;
+
+ case PPC::LBZ:
+ case PPC::LBZ8:
+ case PPC::LD:
+ case PPC::LFD:
+ case PPC::LFS:
+ case PPC::LHA:
+ case PPC::LHA8:
+ case PPC::LHZ:
+ case PPC::LHZ8:
+ case PPC::LWA:
+ case PPC::LWZ:
+ case PPC::LWZ8:
+ FirstOp = 0;
+ break;
+
+ case PPC::STB:
+ case PPC::STB8:
+ case PPC::STD:
+ case PPC::STFD:
+ case PPC::STFS:
+ case PPC::STH:
+ case PPC::STH8:
+ case PPC::STW:
+ case PPC::STW8:
+ FirstOp = 1;
+ break;
+ }
+ // If this is a load or store with a zero offset, we may be able to
+ // fold an add-immediate into the memory operation.
+ if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) ||
+ N->getConstantOperandVal(FirstOp) != 0)
+ continue;
-/// createPPCISelDag - This pass converts a legalized DAG into a
+ SDValue Base = N->getOperand(FirstOp + 1);
+ if (!Base.isMachineOpcode())
+ continue;
+
+ unsigned Flags = 0;
+ bool ReplaceFlags = true;
+
+ // When the feeding operation is an add-immediate of some sort,
+ // determine whether we need to add relocation information to the
+ // target flags on the immediate operand when we fold it into the
+ // load instruction.
+ //
+ // For something like ADDItocL, the relocation information is
+ // inferred from the opcode; when we process it in the AsmPrinter,
+ // we add the necessary relocation there. A load, though, can receive
+ // relocation from various flavors of ADDIxxx, so we need to carry
+ // the relocation information in the target flags.
+ switch (Base.getMachineOpcode()) {
+ default: continue;
+
+ case PPC::ADDI8:
+ case PPC::ADDI8L:
+ case PPC::ADDIL:
+ // In some cases (such as TLS) the relocation information
+ // is already in place on the operand, so copying the operand
+ // is sufficient.
+ ReplaceFlags = false;
+ // For these cases, the immediate may not be divisible by 4, in
+ // which case the fold is illegal for DS-form instructions. (The
+ // other cases provide aligned addresses and are always safe.)
+ if ((StorageOpcode == PPC::LWA ||
+ StorageOpcode == PPC::LD ||
+ StorageOpcode == PPC::STD) &&
+ (!isa<ConstantSDNode>(Base.getOperand(1)) ||
+ Base.getConstantOperandVal(1) % 4 != 0))
+ continue;
+ break;
+ case PPC::ADDIdtprelL:
+ Flags = PPCII::MO_DTPREL16_LO;
+ break;
+ case PPC::ADDItlsldL:
+ Flags = PPCII::MO_TLSLD16_LO;
+ break;
+ case PPC::ADDItocL:
+ Flags = PPCII::MO_TOC16_LO;
+ break;
+ }
+
+ // We found an opportunity. Reverse the operands from the add
+ // immediate and substitute them into the load or store. If
+ // needed, update the target flags for the immediate operand to
+ // reflect the necessary relocation information.
+ DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: ");
+ DEBUG(Base->dump(CurDAG));
+ DEBUG(dbgs() << "\nN: ");
+ DEBUG(N->dump(CurDAG));
+ DEBUG(dbgs() << "\n");
+
+ SDValue ImmOpnd = Base.getOperand(1);
+
+ // If the relocation information isn't already present on the
+ // immediate operand, add it now.
+ if (ReplaceFlags) {
+ if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
+ DebugLoc dl = GA->getDebugLoc();
+ const GlobalValue *GV = GA->getGlobal();
+ ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags);
+ } else if (ConstantPoolSDNode *CP =
+ dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
+ const Constant *C = CP->getConstVal();
+ ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
+ CP->getAlignment(),
+ 0, Flags);
+ }
+ }
+
+ if (FirstOp == 1) // Store
+ (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd,
+ Base.getOperand(0), N->getOperand(3));
+ else // Load
+ (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0),
+ N->getOperand(2));
+
+ // The add-immediate may now be dead, in which case remove it.
+ if (Base.getNode()->use_empty())
+ CurDAG->RemoveDeadNode(Base.getNode());
+ }
+}
+
+
+/// createPPCISelDag - This pass converts a legalized DAG into a
/// PowerPC-specific DAG, ready for instruction scheduling.
///
FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM) {
return new PPCDAGToDAGISel(TM);
}
+static void initializePassOnce(PassRegistry &Registry) {
+ const char *Name = "PowerPC DAG->DAG Pattern Instruction Selection";
+ PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID, 0,
+ false, false);
+ Registry.registerPass(*PI, true);
+}
+
+void llvm::initializePPCDAGToDAGISelPass(PassRegistry &Registry) {
+ CALL_ONCE_INITIALIZATION(initializePassOnce);
+}
+