//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "ppc-codegen"
#include "PPC.h"
#include "MCTargetDesc/PPCPredicates.h"
+#include "PPCMachineFunctionInfo.h"
#include "PPCTargetMachine.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/Module.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;
+#define DEBUG_TYPE "ppc-codegen"
+
// FIXME: Remove this once the bug has been fixed!
cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug",
cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden);
///
class PPCDAGToDAGISel : public SelectionDAGISel {
const PPCTargetMachine &TM;
- const PPCTargetLowering &PPCLowering;
- const PPCSubtarget &PPCSubTarget;
+ const PPCTargetLowering *PPCLowering;
+ const PPCSubtarget *PPCSubTarget;
unsigned GlobalBaseReg;
public:
explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
- : SelectionDAGISel(tm), TM(tm),
- PPCLowering(*TM.getTargetLowering()),
- PPCSubTarget(*TM.getSubtargetImpl()) {
+ : SelectionDAGISel(tm), TM(tm),
+ PPCLowering(TM.getSubtargetImpl()->getTargetLowering()),
+ PPCSubTarget(TM.getSubtargetImpl()) {
initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
}
- virtual bool runOnMachineFunction(MachineFunction &MF) {
+ bool runOnMachineFunction(MachineFunction &MF) override {
// Make sure we re-emit a set of the global base reg if necessary
GlobalBaseReg = 0;
+ PPCLowering = TM.getSubtargetImpl()->getTargetLowering();
+ PPCSubTarget = TM.getSubtargetImpl();
SelectionDAGISel::runOnMachineFunction(MF);
- if (!PPCSubTarget.isSVR4ABI())
+ if (!PPCSubTarget->isSVR4ABI())
InsertVRSaveCode(MF);
return true;
}
- virtual void PostprocessISelDAG();
+ void PostprocessISelDAG() override;
/// getI32Imm - Return a target constant with the specified value, of type
/// i32.
/// getSmallIPtrImm - Return a target constant of pointer type.
inline SDValue getSmallIPtrImm(unsigned Imm) {
- return CurDAG->getTargetConstant(Imm, PPCLowering.getPointerTy());
+ return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy());
}
/// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
/// base register. Return the virtual register that holds this value.
SDNode *getGlobalBaseReg();
+ SDNode *getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset = 0);
+
// 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 *Select(SDNode *N) override;
SDNode *SelectBitfieldInsert(SDNode *N);
/// a base register plus a signed 16-bit displacement [r+imm].
bool SelectAddrImm(SDValue N, SDValue &Disp,
SDValue &Base) {
- return PPCLowering.SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
+ return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
}
/// SelectAddrImmOffs - Return true if the operand is valid for a preinc
/// represented as an indexed [r+r] operation. Returns false if it can
/// be represented by [r+imm], which are preferred.
bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) {
- return PPCLowering.SelectAddressRegReg(N, Base, Index, *CurDAG);
+ 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(SDValue N, SDValue &Base, SDValue &Index) {
- return PPCLowering.SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
+ return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, *CurDAG);
}
/// SelectAddrImmX4 - Returns true if the address N can be represented by
/// a base register plus a signed 16-bit displacement that is a multiple of 4.
/// Suitable for use by STD and friends.
bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) {
- return PPCLowering.SelectAddressRegImm(N, Disp, Base, *CurDAG, true);
+ return PPCLowering->SelectAddressRegImm(N, Disp, Base, *CurDAG, true);
}
// Select an address into a single register.
/// a register. The case of adding a (possibly relocatable) constant to a
/// register can be improved, but it is wrong to substitute Reg+Reg for
/// Reg in an asm, because the load or store opcode would have to change.
- virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
- char ConstraintCode,
- std::vector<SDValue> &OutOps) {
- OutOps.push_back(Op);
+ bool SelectInlineAsmMemoryOperand(const SDValue &Op,
+ char ConstraintCode,
+ std::vector<SDValue> &OutOps) override {
+ // We need to make sure that this one operand does not end up in r0
+ // (because we might end up lowering this as 0(%op)).
+ const TargetRegisterInfo *TRI = TM.getSubtargetImpl()->getRegisterInfo();
+ const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1);
+ SDValue RC = CurDAG->getTargetConstant(TRC->getID(), MVT::i32);
+ SDValue NewOp =
+ SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
+ SDLoc(Op), Op.getValueType(),
+ Op, RC), 0);
+
+ OutOps.push_back(NewOp);
return false;
}
void InsertVRSaveCode(MachineFunction &MF);
- virtual const char *getPassName() const {
+ const char *getPassName() const override {
return "PowerPC DAG->DAG Pattern Instruction Selection";
}
SDNode *SelectSETCC(SDNode *N);
void PeepholePPC64();
- void PeepholdCROps();
+ void PeepholePPC64ZExt();
+ void PeepholeCROps();
+
+ bool AllUsersSelectZero(SDNode *N);
+ void SwapAllSelectUsers(SDNode *N);
};
}
unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
- const TargetInstrInfo &TII = *TM.getInstrInfo();
+ const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
MachineBasicBlock &EntryBB = *Fn.begin();
DebugLoc dl;
// Emit the following code into the entry block:
///
SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
if (!GlobalBaseReg) {
- const TargetInstrInfo &TII = *TM.getInstrInfo();
+ const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
// Insert the set of GlobalBaseReg into the first MBB of the function
MachineBasicBlock &FirstMBB = MF->front();
MachineBasicBlock::iterator MBBI = FirstMBB.begin();
+ const Module *M = MF->getFunction()->getParent();
DebugLoc dl;
- if (PPCLowering.getPointerTy() == MVT::i32) {
- GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
- BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
- BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
+ if (PPCLowering->getPointerTy() == MVT::i32) {
+ if (PPCSubTarget->isTargetELF()) {
+ GlobalBaseReg = PPC::R30;
+ if (M->getPICLevel() == PICLevel::Small) {
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MoveGOTtoLR));
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
+ } else {
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
+ unsigned TempReg = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
+ BuildMI(FirstMBB, MBBI, dl,
+ TII.get(PPC::UpdateGBR)).addReg(GlobalBaseReg)
+ .addReg(TempReg, RegState::Define).addReg(GlobalBaseReg);
+ MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
+ }
+ } else {
+ GlobalBaseReg =
+ RegInfo->createVirtualRegister(&PPC::GPRC_NOR0RegClass);
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR));
+ BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR), GlobalBaseReg);
+ }
} else {
- GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RCRegClass);
+ GlobalBaseReg = RegInfo->createVirtualRegister(&PPC::G8RC_NOX0RegClass);
BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MovePCtoLR8));
BuildMI(FirstMBB, MBBI, dl, TII.get(PPC::MFLR8), GlobalBaseReg);
}
}
return CurDAG->getRegister(GlobalBaseReg,
- PPCLowering.getPointerTy()).getNode();
+ PPCLowering->getPointerTy()).getNode();
}
/// isIntS16Immediate - This method tests to see if the node is either a 32-bit
&& isInt32Immediate(N->getOperand(1).getNode(), Imm);
}
+SDNode *PPCDAGToDAGISel::getFrameIndex(SDNode *SN, SDNode *N, unsigned Offset) {
+ SDLoc dl(SN);
+ int FI = cast<FrameIndexSDNode>(N)->getIndex();
+ SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
+ unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
+ if (SN->hasOneUse())
+ return CurDAG->SelectNodeTo(SN, Opc, N->getValueType(0), TFI,
+ getSmallIPtrImm(Offset));
+ return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
+ getSmallIPtrImm(Offset));
+}
+
bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
if (!Val)
return false;
SDLoc dl(N);
APInt LKZ, LKO, RKZ, RKO;
- CurDAG->ComputeMaskedBits(Op0, LKZ, LKO);
- CurDAG->ComputeMaskedBits(Op1, RKZ, RKO);
+ CurDAG->computeKnownBits(Op0, LKZ, LKO);
+ CurDAG->computeKnownBits(Op1, RKZ, RKO);
unsigned TargetMask = LKZ.getZExtValue();
unsigned InsertMask = RKZ.getZExtValue();
SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value;
}
if (Op1Opc == ISD::AND) {
+ // The AND mask might not be a constant, and we need to make sure that
+ // if we're going to fold the masking with the insert, all bits not
+ // know to be zero in the mask are known to be one.
+ APInt MKZ, MKO;
+ CurDAG->computeKnownBits(Op1.getOperand(1), MKZ, MKO);
+ bool CanFoldMask = InsertMask == MKO.getZExtValue();
+
unsigned SHOpc = Op1.getOperand(0).getOpcode();
- if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) &&
+ if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask &&
isInt32Immediate(Op1.getOperand(0).getOperand(1), Value)) {
- // Note that Value must be in range here (less than 32) because
- // otherwise there would not be any bits set in InsertMask.
+ // Note that Value must be in range here (less than 32) because
+ // otherwise there would not be any bits set in InsertMask.
Op1 = Op1.getOperand(0).getOperand(0);
SH = (SHOpc == ISD::SHL) ? Value : 32 - Value;
}
return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
}
}
- return 0;
+ return nullptr;
}
/// SelectCC - Select a comparison of the specified values with the specified
Opc = PPC::FCMPUS;
} else {
assert(LHS.getValueType() == MVT::f64 && "Unknown vt!");
- Opc = PPC::FCMPUD;
+ Opc = PPCSubTarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD;
}
return SDValue(CurDAG->getMachineNode(Opc, dl, MVT::i32, LHS, RHS), 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");
-}
+static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC,
+ bool HasVSX, bool &Swap, bool &Negate) {
+ Swap = false;
+ Negate = false;
-// 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");
+ if (VecVT.isFloatingPoint()) {
+ /* Handle some cases by swapping input operands. */
+ switch (CC) {
+ case ISD::SETLE: CC = ISD::SETGE; Swap = true; break;
+ case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
+ case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break;
+ case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break;
+ case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
+ case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break;
+ default: break;
+ }
+ /* Handle some cases by negating the result. */
+ switch (CC) {
+ case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
+ case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break;
+ case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break;
+ case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break;
+ default: break;
+ }
+ /* We have instructions implementing the remaining cases. */
+ switch (CC) {
+ case ISD::SETEQ:
+ case ISD::SETOEQ:
+ if (VecVT == MVT::v4f32)
+ return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP;
+ else if (VecVT == MVT::v2f64)
+ return PPC::XVCMPEQDP;
+ break;
+ case ISD::SETGT:
+ case ISD::SETOGT:
+ if (VecVT == MVT::v4f32)
+ return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP;
+ else if (VecVT == MVT::v2f64)
+ return PPC::XVCMPGTDP;
+ break;
+ case ISD::SETGE:
+ case ISD::SETOGE:
+ if (VecVT == MVT::v4f32)
+ return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP;
+ else if (VecVT == MVT::v2f64)
+ return PPC::XVCMPGEDP;
+ break;
+ default:
+ break;
+ }
+ llvm_unreachable("Invalid floating-point vector compare condition");
+ } else {
+ /* Handle some cases by swapping input operands. */
+ switch (CC) {
+ case ISD::SETGE: CC = ISD::SETLE; Swap = true; break;
+ case ISD::SETLT: CC = ISD::SETGT; Swap = true; break;
+ case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break;
+ case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break;
+ default: break;
+ }
+ /* Handle some cases by negating the result. */
+ switch (CC) {
+ case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break;
+ case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break;
+ case ISD::SETLE: CC = ISD::SETGT; Negate = true; break;
+ case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break;
+ default: break;
+ }
+ /* We have instructions implementing the remaining cases. */
+ switch (CC) {
+ case ISD::SETEQ:
+ case ISD::SETUEQ:
+ if (VecVT == MVT::v16i8)
+ return PPC::VCMPEQUB;
+ else if (VecVT == MVT::v8i16)
+ return PPC::VCMPEQUH;
+ else if (VecVT == MVT::v4i32)
+ return PPC::VCMPEQUW;
+ break;
+ case ISD::SETGT:
+ if (VecVT == MVT::v16i8)
+ return PPC::VCMPGTSB;
+ else if (VecVT == MVT::v8i16)
+ return PPC::VCMPGTSH;
+ else if (VecVT == MVT::v4i32)
+ return PPC::VCMPGTSW;
+ break;
+ case ISD::SETUGT:
+ if (VecVT == MVT::v16i8)
+ return PPC::VCMPGTUB;
+ else if (VecVT == MVT::v8i16)
+ return PPC::VCMPGTUH;
+ else if (VecVT == MVT::v4i32)
+ return PPC::VCMPGTUW;
+ break;
+ default:
+ break;
+ }
+ llvm_unreachable("Invalid integer vector compare condition");
}
}
-
SDNode *PPCDAGToDAGISel::SelectSETCC(SDNode *N) {
SDLoc dl(N);
unsigned Imm;
EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
bool isPPC64 = (PtrVT == MVT::i64);
- if (!PPCSubTarget.useCRBits() &&
+ if (!PPCSubTarget->useCRBits() &&
isInt32Immediate(N->getOperand(1), Imm)) {
// We can codegen setcc op, imm very efficiently compared to a brcond.
// Check for those cases here.
case ISD::SETEQ: {
Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) };
- return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
case ISD::SETNE: {
if (isPPC64) break;
}
case ISD::SETLT: {
SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
- return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
case ISD::SETGT: {
SDValue T =
SDValue(CurDAG->getMachineNode(PPC::NEG, dl, MVT::i32, Op), 0);
T = SDValue(CurDAG->getMachineNode(PPC::ANDC, dl, MVT::i32, T, Op), 0);
SDValue Ops[] = { T, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
- return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
}
} else if (Imm == ~0U) { // setcc op, -1
SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
Op), 0);
SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
- return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
case ISD::SETGT: {
SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
// 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 Swap, Negate;
+ unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC,
+ PPCSubTarget->hasVSX(), Swap, Negate);
+ if (Swap)
+ std::swap(LHS, RHS);
+
+ if (Negate) {
+ SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
+ return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR :
+ PPC::VNOR,
+ VecVT, VCmp, VCmp);
}
+
+ return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
}
- if (PPCSubTarget.useCRBits())
- return 0;
+ if (PPCSubTarget->useCRBits())
+ return nullptr;
bool Inv;
unsigned Idx = getCRIdxForSetCC(CC, Inv);
// Force the ccreg into CR7.
SDValue CR7Reg = CurDAG->getRegister(PPC::CR7, MVT::i32);
- SDValue InFlag(0, 0); // Null incoming flag value.
+ SDValue InFlag(nullptr, 0); // Null incoming flag value.
CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
InFlag).getValue(1);
SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
getI32Imm(31), getI32Imm(31) };
if (!Inv)
- return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
// Get the specified bit.
SDValue Tmp =
SDLoc dl(N);
if (N->isMachineOpcode()) {
N->setNodeId(-1);
- return NULL; // Already selected.
+ return nullptr; // Already selected.
}
+ // In case any misguided DAG-level optimizations form an ADD with a
+ // TargetConstant operand, crash here instead of miscompiling (by selecting
+ // an r+r add instead of some kind of r+i add).
+ if (N->getOpcode() == ISD::ADD &&
+ N->getOperand(1).getOpcode() == ISD::TargetConstant)
+ llvm_unreachable("Invalid ADD with TargetConstant operand");
+
switch (N->getOpcode()) {
default: break;
case PPCISD::GlobalBaseReg:
return getGlobalBaseReg();
- case ISD::FrameIndex: {
- int FI = cast<FrameIndexSDNode>(N)->getIndex();
- SDValue TFI = CurDAG->getTargetFrameIndex(FI, N->getValueType(0));
- unsigned Opc = N->getValueType(0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8;
- if (N->hasOneUse())
- return CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), TFI,
- getSmallIPtrImm(0));
- return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
- getSmallIPtrImm(0));
- }
+ case ISD::FrameIndex:
+ return getFrameIndex(N, N);
case PPCISD::MFOCRF: {
SDValue InFlag = N->getOperand(1);
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
- // 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;
- if (isInt32Immediate(N->getOperand(1), Imm)) {
- SDValue N0 = N->getOperand(0);
- if ((signed)Imm > 0 && isPowerOf2_32(Imm)) {
- SDNode *Op =
- CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
- N0, getI32Imm(Log2_32(Imm)));
- 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::Glue,
- N0, getI32Imm(Log2_32(-Imm)));
- SDValue PT =
- SDValue(CurDAG->getMachineNode(PPC::ADDZE, dl, MVT::i32,
- SDValue(Op, 0), SDValue(Op, 1)),
- 0);
- return CurDAG->SelectNodeTo(N, PPC::NEG, MVT::i32, PT);
- }
- }
+ case PPCISD::READ_TIME_BASE: {
+ return CurDAG->getMachineNode(PPC::ReadTB, dl, MVT::i32, MVT::i32,
+ MVT::Other, N->getOperand(0));
+ }
- // Other cases are autogenerated.
- break;
+ case PPCISD::SRA_ADDZE: {
+ SDValue N0 = N->getOperand(0);
+ SDValue ShiftAmt =
+ CurDAG->getTargetConstant(*cast<ConstantSDNode>(N->getOperand(1))->
+ getConstantIntValue(), N->getValueType(0));
+ if (N->getValueType(0) == MVT::i64) {
+ SDNode *Op =
+ CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue,
+ N0, ShiftAmt);
+ return CurDAG->SelectNodeTo(N, PPC::ADDZE8, MVT::i64,
+ SDValue(Op, 0), SDValue(Op, 1));
+ } else {
+ assert(N->getValueType(0) == MVT::i32 &&
+ "Expecting i64 or i32 in PPCISD::SRA_ADDZE");
+ SDNode *Op =
+ CurDAG->getMachineNode(PPC::SRAWI, dl, MVT::i32, MVT::Glue,
+ N0, ShiftAmt);
+ return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
+ SDValue(Op, 0), SDValue(Op, 1));
+ }
}
case ISD::LOAD: {
SDValue Base = LD->getBasePtr();
SDValue Ops[] = { Offset, Base, Chain };
return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
- PPCLowering.getPointerTy(),
+ PPCLowering->getPointerTy(),
MVT::Other, Ops);
} else {
unsigned Opcode;
SDValue Base = LD->getBasePtr();
SDValue Ops[] = { Base, Offset, Chain };
return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
- PPCLowering.getPointerTy(),
+ PPCLowering->getPointerTy(),
MVT::Other, Ops);
}
}
isRotateAndMask(N->getOperand(0).getNode(), Imm, false, SH, MB, ME)) {
SDValue Val = N->getOperand(0).getOperand(0);
SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB), getI32Imm(ME) };
- return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
// 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
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);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
// If this is a 64-bit zero-extension mask, emit rldicl.
if (isInt64Immediate(N->getOperand(1).getNode(), Imm64) &&
}
SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) };
- return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops, 3);
+ return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops);
}
// AND X, 0 -> 0, not "rlwinm 32".
if (isInt32Immediate(N->getOperand(1), Imm) && (Imm == 0)) {
ReplaceUses(SDValue(N, 0), N->getOperand(1));
- return NULL;
+ return nullptr;
}
// ISD::OR doesn't get all the bitfield insertion fun.
// (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) is a bitfield insert
// Other cases are autogenerated.
break;
}
- case ISD::OR:
+ case ISD::OR: {
if (N->getValueType(0) == MVT::i32)
if (SDNode *I = SelectBitfieldInsert(N))
return I;
+ short Imm;
+ if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
+ isIntS16Immediate(N->getOperand(1), Imm)) {
+ APInt LHSKnownZero, LHSKnownOne;
+ CurDAG->computeKnownBits(N->getOperand(0), LHSKnownZero, LHSKnownOne);
+
+ // If this is equivalent to an add, then we can fold it with the
+ // FrameIndex calculation.
+ if ((LHSKnownZero.getZExtValue()|~(uint64_t)Imm) == ~0ULL)
+ return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
+ }
+
// Other cases are autogenerated.
break;
+ }
+ case ISD::ADD: {
+ short Imm;
+ if (N->getOperand(0)->getOpcode() == ISD::FrameIndex &&
+ isIntS16Immediate(N->getOperand(1), Imm))
+ return getFrameIndex(N, N->getOperand(0).getNode(), (int)Imm);
+
+ break;
+ }
case ISD::SHL: {
unsigned Imm, SH, MB, ME;
if (isOpcWithIntImmediate(N->getOperand(0).getNode(), ISD::AND, Imm) &&
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);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
// Other cases are autogenerated.
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);
+ return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
// Other cases are autogenerated.
bool isPPC64 = (PtrVT == MVT::i64);
// If this is a select of i1 operands, we'll pattern match it.
- if (PPCSubTarget.useCRBits() &&
+ if (PPCSubTarget->useCRBits() &&
N->getOperand(0).getValueType() == MVT::i1)
break;
else if (N->getValueType(0) == MVT::f32)
SelectCCOp = PPC::SELECT_CC_F4;
else if (N->getValueType(0) == MVT::f64)
- SelectCCOp = PPC::SELECT_CC_F8;
+ if (PPCSubTarget->hasVSX())
+ SelectCCOp = PPC::SELECT_CC_VSFRC;
+ else
+ SelectCCOp = PPC::SELECT_CC_F8;
+ else if (N->getValueType(0) == MVT::v2f64 ||
+ N->getValueType(0) == MVT::v2i64)
+ SelectCCOp = PPC::SELECT_CC_VSRC;
else
SelectCCOp = PPC::SELECT_CC_VRRC;
SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
getI32Imm(BROpc) };
- return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops, 4);
+ return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops);
}
+ case ISD::VSELECT:
+ if (PPCSubTarget->hasVSX()) {
+ SDValue Ops[] = { N->getOperand(2), N->getOperand(1), N->getOperand(0) };
+ return CurDAG->SelectNodeTo(N, PPC::XXSEL, N->getValueType(0), Ops);
+ }
+
+ break;
+ case ISD::VECTOR_SHUFFLE:
+ if (PPCSubTarget->hasVSX() && (N->getValueType(0) == MVT::v2f64 ||
+ N->getValueType(0) == MVT::v2i64)) {
+ ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
+
+ SDValue Op1 = N->getOperand(SVN->getMaskElt(0) < 2 ? 0 : 1),
+ Op2 = N->getOperand(SVN->getMaskElt(1) < 2 ? 0 : 1);
+ unsigned DM[2];
+
+ for (int i = 0; i < 2; ++i)
+ if (SVN->getMaskElt(i) <= 0 || SVN->getMaskElt(i) == 2)
+ DM[i] = 0;
+ else
+ DM[i] = 1;
+
+ // For little endian, we must swap the input operands and adjust
+ // the mask elements (reverse and invert them).
+ if (PPCSubTarget->isLittleEndian()) {
+ std::swap(Op1, Op2);
+ unsigned tmp = DM[0];
+ DM[0] = 1 - DM[1];
+ DM[1] = 1 - tmp;
+ }
+
+ SDValue DMV = CurDAG->getTargetConstant(DM[1] | (DM[0] << 1), MVT::i32);
+
+ if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 &&
+ Op1.getOpcode() == ISD::SCALAR_TO_VECTOR &&
+ isa<LoadSDNode>(Op1.getOperand(0))) {
+ LoadSDNode *LD = cast<LoadSDNode>(Op1.getOperand(0));
+ SDValue Base, Offset;
+
+ if (LD->isUnindexed() &&
+ SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) {
+ SDValue Chain = LD->getChain();
+ SDValue Ops[] = { Base, Offset, Chain };
+ return CurDAG->SelectNodeTo(N, PPC::LXVDSX,
+ N->getValueType(0), Ops);
+ }
+ }
+
+ SDValue Ops[] = { Op1, Op2, DMV };
+ return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops);
+ }
+
+ break;
case PPCISD::BDNZ:
case PPCISD::BDZ: {
- bool IsPPC64 = PPCSubTarget.isPPC64();
+ bool IsPPC64 = PPCSubTarget->isPPC64();
SDValue Ops[] = { N->getOperand(1), N->getOperand(0) };
return CurDAG->SelectNodeTo(N, N->getOpcode() == PPCISD::BDNZ ?
(IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) :
(IsPPC64 ? PPC::BDZ8 : PPC::BDZ),
- MVT::Other, Ops, 2);
+ MVT::Other, Ops);
}
case PPCISD::COND_BRANCH: {
// Op #0 is the Chain.
getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue());
SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
N->getOperand(0), N->getOperand(4) };
- return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops, 5);
+ return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
}
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(PCC), CondCode,
N->getOperand(4), N->getOperand(0) };
- return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops, 4);
+ return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
}
case ISD::BRIND: {
// FIXME: Should custom lower this.
return CurDAG->SelectNodeTo(N, Reg, MVT::Other, Chain);
}
case PPCISD::TOC_ENTRY: {
- assert (PPCSubTarget.isPPC64() && "Only supported for 64-bit ABI");
+ assert ((PPCSubTarget->isPPC64() || PPCSubTarget->isSVR4ABI()) &&
+ "Only supported for 64-bit ABI and 32-bit SVR4");
+ if (PPCSubTarget->isSVR4ABI() && !PPCSubTarget->isPPC64()) {
+ SDValue GA = N->getOperand(0);
+ return CurDAG->getMachineNode(PPC::LWZtoc, dl, MVT::i32, GA,
+ N->getOperand(1));
+ }
// 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.
+ // selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA.
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:
+ // The first source operand is a TargetGlobalAddress or a TargetJumpTable.
+ // If it is an externally defined symbol, a symbol with common linkage,
+ // a non-local 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>)
SDNode *Tmp = CurDAG->getMachineNode(PPC::ADDIStocHA, dl, MVT::i64,
TOCbase, GA);
- if (isa<JumpTableSDNode>(GA) || CModel == CodeModel::Large)
+ if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(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())
+ if ((GValue->getType()->getElementType()->isFunctionTy() &&
+ (GValue->isDeclaration() || GValue->isWeakForLinker())) ||
+ GValue->isDeclaration() || GValue->hasCommonLinkage() ||
+ GValue->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::PPC32_PICGOT: {
+ // Generate a PIC-safe GOT reference.
+ assert(!PPCSubTarget->isPPC64() && PPCSubTarget->isSVR4ABI() &&
+ "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4");
+ return CurDAG->SelectNodeTo(N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(), MVT::i32);
+ }
case PPCISD::VADD_SPLAT: {
// This expands into one of three sequences, depending on whether
// the first operand is odd or even, positive or negative.
return;
PeepholePPC64();
- PeepholdCROps();
+ PeepholeCROps();
+ PeepholePPC64ZExt();
+}
+
+// Check if all users of this node will become isel where the second operand
+// is the constant zero. If this is so, and if we can negate the condition,
+// then we can flip the true and false operands. This will allow the zero to
+// be folded with the isel so that we don't need to materialize a register
+// containing zero.
+bool PPCDAGToDAGISel::AllUsersSelectZero(SDNode *N) {
+ // If we're not using isel, then this does not matter.
+ if (!PPCSubTarget->hasISEL())
+ return false;
+
+ for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ if (!User->isMachineOpcode())
+ return false;
+ if (User->getMachineOpcode() != PPC::SELECT_I4 &&
+ User->getMachineOpcode() != PPC::SELECT_I8)
+ return false;
+
+ SDNode *Op2 = User->getOperand(2).getNode();
+ if (!Op2->isMachineOpcode())
+ return false;
+
+ if (Op2->getMachineOpcode() != PPC::LI &&
+ Op2->getMachineOpcode() != PPC::LI8)
+ return false;
+
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op2->getOperand(0));
+ if (!C)
+ return false;
+
+ if (!C->isNullValue())
+ return false;
+ }
+
+ return true;
}
-void PPCDAGToDAGISel::PeepholdCROps() {
+void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) {
+ SmallVector<SDNode *, 4> ToReplace;
+ for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
+ UI != UE; ++UI) {
+ SDNode *User = *UI;
+ assert((User->getMachineOpcode() == PPC::SELECT_I4 ||
+ User->getMachineOpcode() == PPC::SELECT_I8) &&
+ "Must have all select users");
+ ToReplace.push_back(User);
+ }
+
+ for (SmallVector<SDNode *, 4>::iterator UI = ToReplace.begin(),
+ UE = ToReplace.end(); UI != UE; ++UI) {
+ SDNode *User = *UI;
+ SDNode *ResNode =
+ CurDAG->getMachineNode(User->getMachineOpcode(), SDLoc(User),
+ User->getValueType(0), User->getOperand(0),
+ User->getOperand(2),
+ User->getOperand(1));
+
+ DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
+ DEBUG(User->dump(CurDAG));
+ DEBUG(dbgs() << "\nNew: ");
+ DEBUG(ResNode->dump(CurDAG));
+ DEBUG(dbgs() << "\n");
+
+ ReplaceUses(User, ResNode);
+ }
+}
+
+void PPCDAGToDAGISel::PeepholeCROps() {
bool IsModified;
do {
IsModified = false;
case PPC::SELECT_I8:
case PPC::SELECT_F4:
case PPC::SELECT_F8:
- case PPC::SELECT_VRRC: {
+ case PPC::SELECT_VRRC:
+ case PPC::SELECT_VSFRC:
+ case PPC::SELECT_VSRC: {
SDValue Op = MachineNode->getOperand(0);
if (Op.isMachineOpcode()) {
if (Op.getMachineOpcode() == PPC::CRSET)
break;
}
+ bool SelectSwap = false;
switch (Opcode) {
default: break;
case PPC::CRAND:
MVT::i1, MachineNode->getOperand(0),
MachineNode->getOperand(1).
getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(1)),
+ SelectSwap = true;
break;
case PPC::CRNAND:
if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
MVT::i1, MachineNode->getOperand(1).
getOperand(0),
MachineNode->getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(1)),
+ SelectSwap = true;
break;
case PPC::CROR:
if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
MVT::i1, MachineNode->getOperand(0),
MachineNode->getOperand(1).
getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(1)),
+ SelectSwap = true;
break;
case PPC::CRXOR:
if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
MVT::i1, MachineNode->getOperand(0),
MachineNode->getOperand(1).
getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(1)),
+ SelectSwap = true;
break;
case PPC::CRNOR:
if (Op1Set || Op2Set)
MVT::i1, MachineNode->getOperand(1).
getOperand(0),
MachineNode->getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(1)),
+ SelectSwap = true;
break;
case PPC::CREQV:
if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
MVT::i1, MachineNode->getOperand(0),
MachineNode->getOperand(1).
getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(1)),
+ SelectSwap = true;
break;
case PPC::CRANDC:
if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
MVT::i1, MachineNode->getOperand(0),
MachineNode->getOperand(1).
getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(0)),
+ SelectSwap = true;
break;
case PPC::CRORC:
if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
MVT::i1, MachineNode->getOperand(0),
MachineNode->getOperand(1).
getOperand(0));
+ else if (AllUsersSelectZero(MachineNode))
+ ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(0)),
+ SelectSwap = true;
break;
case PPC::SELECT_I4:
case PPC::SELECT_I8:
case PPC::SELECT_F4:
case PPC::SELECT_F8:
case PPC::SELECT_VRRC:
+ case PPC::SELECT_VSFRC:
+ case PPC::SELECT_VSRC:
if (Op1Set)
ResNode = MachineNode->getOperand(1).getNode();
else if (Op1Unset)
break;
}
+ // If we're inverting this node because it is used only by selects that
+ // we'd like to swap, then swap the selects before the node replacement.
+ if (SelectSwap)
+ SwapAllSelectUsers(MachineNode);
+
if (ResNode != MachineNode) {
DEBUG(dbgs() << "CR Peephole replacing:\nOld: ");
DEBUG(MachineNode->dump(CurDAG));
} while (IsModified);
}
+// Gather the set of 32-bit operations that are known to have their
+// higher-order 32 bits zero, where ToPromote contains all such operations.
+static bool PeepholePPC64ZExtGather(SDValue Op32,
+ SmallPtrSetImpl<SDNode *> &ToPromote) {
+ if (!Op32.isMachineOpcode())
+ return false;
+
+ // First, check for the "frontier" instructions (those that will clear the
+ // higher-order 32 bits.
+
+ // For RLWINM and RLWNM, we need to make sure that the mask does not wrap
+ // around. If it does not, then these instructions will clear the
+ // higher-order bits.
+ if ((Op32.getMachineOpcode() == PPC::RLWINM ||
+ Op32.getMachineOpcode() == PPC::RLWNM) &&
+ Op32.getConstantOperandVal(2) <= Op32.getConstantOperandVal(3)) {
+ ToPromote.insert(Op32.getNode());
+ return true;
+ }
+
+ // SLW and SRW always clear the higher-order bits.
+ if (Op32.getMachineOpcode() == PPC::SLW ||
+ Op32.getMachineOpcode() == PPC::SRW) {
+ ToPromote.insert(Op32.getNode());
+ return true;
+ }
+
+ // For LI and LIS, we need the immediate to be positive (so that it is not
+ // sign extended).
+ if (Op32.getMachineOpcode() == PPC::LI ||
+ Op32.getMachineOpcode() == PPC::LIS) {
+ if (!isUInt<15>(Op32.getConstantOperandVal(0)))
+ return false;
+
+ ToPromote.insert(Op32.getNode());
+ return true;
+ }
+
+ // Next, check for those instructions we can look through.
+
+ // Assuming the mask does not wrap around, then the higher-order bits are
+ // taken directly from the first operand.
+ if (Op32.getMachineOpcode() == PPC::RLWIMI &&
+ Op32.getConstantOperandVal(3) <= Op32.getConstantOperandVal(4)) {
+ SmallPtrSet<SDNode *, 16> ToPromote1;
+ if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
+ return false;
+
+ ToPromote.insert(Op32.getNode());
+ ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
+ return true;
+ }
+
+ // For OR, the higher-order bits are zero if that is true for both operands.
+ // For SELECT_I4, the same is true (but the relevant operand numbers are
+ // shifted by 1).
+ if (Op32.getMachineOpcode() == PPC::OR ||
+ Op32.getMachineOpcode() == PPC::SELECT_I4) {
+ unsigned B = Op32.getMachineOpcode() == PPC::SELECT_I4 ? 1 : 0;
+ SmallPtrSet<SDNode *, 16> ToPromote1;
+ if (!PeepholePPC64ZExtGather(Op32.getOperand(B+0), ToPromote1))
+ return false;
+ if (!PeepholePPC64ZExtGather(Op32.getOperand(B+1), ToPromote1))
+ return false;
+
+ ToPromote.insert(Op32.getNode());
+ ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
+ return true;
+ }
+
+ // For ORI and ORIS, we need the higher-order bits of the first operand to be
+ // zero, and also for the constant to be positive (so that it is not sign
+ // extended).
+ if (Op32.getMachineOpcode() == PPC::ORI ||
+ Op32.getMachineOpcode() == PPC::ORIS) {
+ SmallPtrSet<SDNode *, 16> ToPromote1;
+ if (!PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1))
+ return false;
+ if (!isUInt<15>(Op32.getConstantOperandVal(1)))
+ return false;
+
+ ToPromote.insert(Op32.getNode());
+ ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
+ return true;
+ }
+
+ // The higher-order bits of AND are zero if that is true for at least one of
+ // the operands.
+ if (Op32.getMachineOpcode() == PPC::AND) {
+ SmallPtrSet<SDNode *, 16> ToPromote1, ToPromote2;
+ bool Op0OK =
+ PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
+ bool Op1OK =
+ PeepholePPC64ZExtGather(Op32.getOperand(1), ToPromote2);
+ if (!Op0OK && !Op1OK)
+ return false;
+
+ ToPromote.insert(Op32.getNode());
+
+ if (Op0OK)
+ ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
+
+ if (Op1OK)
+ ToPromote.insert(ToPromote2.begin(), ToPromote2.end());
+
+ return true;
+ }
+
+ // For ANDI and ANDIS, the higher-order bits are zero if either that is true
+ // of the first operand, or if the second operand is positive (so that it is
+ // not sign extended).
+ if (Op32.getMachineOpcode() == PPC::ANDIo ||
+ Op32.getMachineOpcode() == PPC::ANDISo) {
+ SmallPtrSet<SDNode *, 16> ToPromote1;
+ bool Op0OK =
+ PeepholePPC64ZExtGather(Op32.getOperand(0), ToPromote1);
+ bool Op1OK = isUInt<15>(Op32.getConstantOperandVal(1));
+ if (!Op0OK && !Op1OK)
+ return false;
+
+ ToPromote.insert(Op32.getNode());
+
+ if (Op0OK)
+ ToPromote.insert(ToPromote1.begin(), ToPromote1.end());
+
+ return true;
+ }
+
+ return false;
+}
+
+void PPCDAGToDAGISel::PeepholePPC64ZExt() {
+ if (!PPCSubTarget->isPPC64())
+ return;
+
+ // When we zero-extend from i32 to i64, we use a pattern like this:
+ // def : Pat<(i64 (zext i32:$in)),
+ // (RLDICL (INSERT_SUBREG (i64 (IMPLICIT_DEF)), $in, sub_32),
+ // 0, 32)>;
+ // There are several 32-bit shift/rotate instructions, however, that will
+ // clear the higher-order bits of their output, rendering the RLDICL
+ // unnecessary. When that happens, we remove it here, and redefine the
+ // relevant 32-bit operation to be a 64-bit operation.
+
+ SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
+ ++Position;
+
+ bool MadeChange = false;
+ while (Position != CurDAG->allnodes_begin()) {
+ SDNode *N = --Position;
+ // Skip dead nodes and any non-machine opcodes.
+ if (N->use_empty() || !N->isMachineOpcode())
+ continue;
+
+ if (N->getMachineOpcode() != PPC::RLDICL)
+ continue;
+
+ if (N->getConstantOperandVal(1) != 0 ||
+ N->getConstantOperandVal(2) != 32)
+ continue;
+
+ SDValue ISR = N->getOperand(0);
+ if (!ISR.isMachineOpcode() ||
+ ISR.getMachineOpcode() != TargetOpcode::INSERT_SUBREG)
+ continue;
+
+ if (!ISR.hasOneUse())
+ continue;
+
+ if (ISR.getConstantOperandVal(2) != PPC::sub_32)
+ continue;
+
+ SDValue IDef = ISR.getOperand(0);
+ if (!IDef.isMachineOpcode() ||
+ IDef.getMachineOpcode() != TargetOpcode::IMPLICIT_DEF)
+ continue;
+
+ // We now know that we're looking at a canonical i32 -> i64 zext. See if we
+ // can get rid of it.
+
+ SDValue Op32 = ISR->getOperand(1);
+ if (!Op32.isMachineOpcode())
+ continue;
+
+ // There are some 32-bit instructions that always clear the high-order 32
+ // bits, there are also some instructions (like AND) that we can look
+ // through.
+ SmallPtrSet<SDNode *, 16> ToPromote;
+ if (!PeepholePPC64ZExtGather(Op32, ToPromote))
+ continue;
+
+ // If the ToPromote set contains nodes that have uses outside of the set
+ // (except for the original INSERT_SUBREG), then abort the transformation.
+ bool OutsideUse = false;
+ for (SDNode *PN : ToPromote) {
+ for (SDNode *UN : PN->uses()) {
+ if (!ToPromote.count(UN) && UN != ISR.getNode()) {
+ OutsideUse = true;
+ break;
+ }
+ }
+
+ if (OutsideUse)
+ break;
+ }
+ if (OutsideUse)
+ continue;
+
+ MadeChange = true;
+
+ // We now know that this zero extension can be removed by promoting to
+ // nodes in ToPromote to 64-bit operations, where for operations in the
+ // frontier of the set, we need to insert INSERT_SUBREGs for their
+ // operands.
+ for (SDNode *PN : ToPromote) {
+ unsigned NewOpcode;
+ switch (PN->getMachineOpcode()) {
+ default:
+ llvm_unreachable("Don't know the 64-bit variant of this instruction");
+ case PPC::RLWINM: NewOpcode = PPC::RLWINM8; break;
+ case PPC::RLWNM: NewOpcode = PPC::RLWNM8; break;
+ case PPC::SLW: NewOpcode = PPC::SLW8; break;
+ case PPC::SRW: NewOpcode = PPC::SRW8; break;
+ case PPC::LI: NewOpcode = PPC::LI8; break;
+ case PPC::LIS: NewOpcode = PPC::LIS8; break;
+ case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break;
+ case PPC::OR: NewOpcode = PPC::OR8; break;
+ case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break;
+ case PPC::ORI: NewOpcode = PPC::ORI8; break;
+ case PPC::ORIS: NewOpcode = PPC::ORIS8; break;
+ case PPC::AND: NewOpcode = PPC::AND8; break;
+ case PPC::ANDIo: NewOpcode = PPC::ANDIo8; break;
+ case PPC::ANDISo: NewOpcode = PPC::ANDISo8; break;
+ }
+
+ // Note: During the replacement process, the nodes will be in an
+ // inconsistent state (some instructions will have operands with values
+ // of the wrong type). Once done, however, everything should be right
+ // again.
+
+ SmallVector<SDValue, 4> Ops;
+ for (const SDValue &V : PN->ops()) {
+ if (!ToPromote.count(V.getNode()) && V.getValueType() == MVT::i32 &&
+ !isa<ConstantSDNode>(V)) {
+ SDValue ReplOpOps[] = { ISR.getOperand(0), V, ISR.getOperand(2) };
+ SDNode *ReplOp =
+ CurDAG->getMachineNode(TargetOpcode::INSERT_SUBREG, SDLoc(V),
+ ISR.getNode()->getVTList(), ReplOpOps);
+ Ops.push_back(SDValue(ReplOp, 0));
+ } else {
+ Ops.push_back(V);
+ }
+ }
+
+ // Because all to-be-promoted nodes only have users that are other
+ // promoted nodes (or the original INSERT_SUBREG), we can safely replace
+ // the i32 result value type with i64.
+
+ SmallVector<EVT, 2> NewVTs;
+ SDVTList VTs = PN->getVTList();
+ for (unsigned i = 0, ie = VTs.NumVTs; i != ie; ++i)
+ if (VTs.VTs[i] == MVT::i32)
+ NewVTs.push_back(MVT::i64);
+ else
+ NewVTs.push_back(VTs.VTs[i]);
+
+ DEBUG(dbgs() << "PPC64 ZExt Peephole morphing:\nOld: ");
+ DEBUG(PN->dump(CurDAG));
+
+ CurDAG->SelectNodeTo(PN, NewOpcode, CurDAG->getVTList(NewVTs), Ops);
+
+ DEBUG(dbgs() << "\nNew: ");
+ DEBUG(PN->dump(CurDAG));
+ DEBUG(dbgs() << "\n");
+ }
+
+ // Now we replace the original zero extend and its associated INSERT_SUBREG
+ // with the value feeding the INSERT_SUBREG (which has now been promoted to
+ // return an i64).
+
+ DEBUG(dbgs() << "PPC64 ZExt Peephole replacing:\nOld: ");
+ DEBUG(N->dump(CurDAG));
+ DEBUG(dbgs() << "\nNew: ");
+ DEBUG(Op32.getNode()->dump(CurDAG));
+ DEBUG(dbgs() << "\n");
+
+ ReplaceUses(N, Op32.getNode());
+ }
+
+ if (MadeChange)
+ CurDAG->RemoveDeadNodes();
+}
+
void PPCDAGToDAGISel::PeepholePPC64() {
// These optimizations are currently supported only for 64-bit SVR4.
- if (PPCSubTarget.isDarwin() || !PPCSubTarget.isPPC64())
+ if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64())
return;
SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
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);
+ PassInfo *PI = new PassInfo(Name, "ppc-codegen", &SelectionDAGISel::ID,
+ nullptr, false, false);
Registry.registerPass(*PI, true);
}