#define DEBUG_TYPE "ppc-codegen"
#include "PPC.h"
-#include "PPCTargetMachine.h"
#include "MCTargetDesc/PPCPredicates.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "PPCTargetMachine.h"
#include "llvm/CodeGen/MachineFunction.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/CommandLine.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;
+// 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);
+
+namespace llvm {
+ void initializePPCDAGToDAGISelPass(PassRegistry&);
+}
+
namespace {
//===--------------------------------------------------------------------===//
/// PPCDAGToDAGISel - PPC specific code to select PPC machine
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) {
/// SelectCC - Select a comparison of the specified values with the
/// specified condition code, returning the CR# of the expression.
- SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, DebugLoc dl);
+ SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, SDLoc dl);
/// SelectAddrImm - Returns true if the address N can be represented by
/// 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);
+ return PPCLowering.SelectAddressRegImm(N, Disp, Base, *CurDAG, false);
}
/// SelectAddrImmOffs - Return true if the operand is valid for a preinc
- /// immediate field. Because preinc imms have already been validated, just
- /// accept it.
+ /// 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 (isa<ConstantSDNode>(N)) {
+ if (N.getOpcode() == ISD::TargetConstant ||
+ N.getOpcode() == ISD::TargetGlobalAddress) {
Out = N;
return true;
}
return false;
}
- /// SelectAddrIdxOffs - Return true if the operand is valid for a preinc
- /// index field. Because preinc imms have already been validated, just
- /// accept it.
- bool SelectAddrIdxOffs(SDValue N, SDValue &Out) const {
- Out = N;
- return true;
- }
-
/// 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.
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(SDValue N, SDValue &Disp, SDValue &Base) {
- return PPCLowering.SelectAddressRegImmShift(N, Disp, Base, *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);
+ }
+
+ // Select an address into a single register.
+ bool SelectAddr(SDValue N, SDValue &Base) {
+ Base = N;
+ return true;
}
/// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
private:
SDNode *SelectSETCC(SDNode *N);
+
+ void PeepholePPC64();
+ void PeepholdCROps();
+
+ bool AllUsersSelectZero(SDNode *N);
+ void SwapAllSelectUsers(SDNode *N);
};
}
}
bool PPCDAGToDAGISel::isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME) {
+ if (!Val)
+ return false;
+
if (isShiftedMask_32(Val)) {
// look for the first non-zero bit
- MB = CountLeadingZeros_32(Val);
+ MB = countLeadingZeros(Val);
// look for the first zero bit after the run of ones
- ME = CountLeadingZeros_32((Val - 1) ^ Val);
+ ME = countLeadingZeros((Val - 1) ^ Val);
return true;
} else {
Val = ~Val; // invert mask
if (isShiftedMask_32(Val)) {
// effectively look for the first zero bit
- ME = CountLeadingZeros_32(Val) - 1;
+ ME = countLeadingZeros(Val) - 1;
// effectively look for the first one bit after the run of zeros
- MB = CountLeadingZeros_32((Val - 1) ^ Val) + 1;
+ MB = countLeadingZeros((Val - 1) ^ Val) + 1;
return true;
}
}
SDNode *PPCDAGToDAGISel::SelectBitfieldInsert(SDNode *N) {
SDValue Op0 = N->getOperand(0);
SDValue Op1 = N->getOperand(1);
- DebugLoc dl = N->getDebugLoc();
+ SDLoc dl(N);
APInt LKZ, LKO, RKZ, RKO;
CurDAG->ComputeMaskedBits(Op0, LKZ, LKO);
}
unsigned MB, ME;
- if (InsertMask && isRunOfOnes(InsertMask, MB, ME)) {
+ if (isRunOfOnes(InsertMask, MB, ME)) {
SDValue Tmp1, Tmp2;
if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) &&
unsigned SHOpc = Op1.getOperand(0).getOpcode();
if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) &&
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.
Op1 = Op1.getOperand(0).getOperand(0);
SH = (SHOpc == ISD::SHL) ? Value : 32 - Value;
- } else {
- Op1 = Op1.getOperand(0);
}
}
SH &= 31;
SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB),
getI32Imm(ME) };
- return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops, 5);
+ return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
}
}
return 0;
/// SelectCC - Select a comparison of the specified values with the specified
/// condition code, returning the CR# of the expression.
SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS,
- ISD::CondCode CC, DebugLoc dl) {
+ ISD::CondCode CC, SDLoc dl) {
// Always select the LHS.
unsigned Opc;
/// getCRIdxForSetCC - Return the index of the condition register field
/// associated with the SetCC condition, and whether or not the field is
/// treated as inverted. That is, lt = 0; ge = 0 inverted.
-///
-/// If this returns with Other != -1, then the returned comparison is an or of
-/// two simpler comparisons. In this case, Invert is guaranteed to be false.
-static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert, int &Other) {
+static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) {
Invert = false;
- Other = -1;
switch (CC) {
default: llvm_unreachable("Unknown condition!");
case ISD::SETOLT:
}
}
+// 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();
+ SDLoc dl(N);
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)) {
+ 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.
// setcc op, 0
}
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),
0);
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");
+ }
+ }
+
+ if (PPCSubTarget.useCRBits())
+ return 0;
+
bool Inv;
- int OtherCondIdx;
- unsigned Idx = getCRIdxForSetCC(CC, Inv, OtherCondIdx);
- SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
+ unsigned Idx = getCRIdxForSetCC(CC, Inv);
+ SDValue CCReg = SelectCC(LHS, RHS, CC, dl);
SDValue IntCR;
// Force the ccreg into CR7.
CCReg = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, CR7Reg, CCReg,
InFlag).getValue(1);
- if (PPCSubTarget.hasMFOCRF() && OtherCondIdx == -1)
- IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
- CCReg), 0);
- else
- IntCR = SDValue(CurDAG->getMachineNode(PPC::MFCRpseud, dl, MVT::i32,
- CR7Reg, CCReg), 0);
+ IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
+ CCReg), 0);
SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
getI32Imm(31), getI32Imm(31) };
- if (OtherCondIdx == -1 && !Inv)
+ if (!Inv)
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops, 4);
// Get the specified bit.
SDValue Tmp =
- SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4), 0);
- if (Inv) {
- assert(OtherCondIdx == -1 && "Can't have split plus negation");
- return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
- }
-
- // Otherwise, we have to turn an operation like SETONE -> SETOLT | SETOGT.
- // We already got the bit for the first part of the comparison (e.g. SETULE).
-
- // Get the other bit of the comparison.
- Ops[1] = getI32Imm((32-(3-OtherCondIdx)) & 31);
- SDValue OtherCond =
- SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops, 4), 0);
-
- return CurDAG->SelectNodeTo(N, PPC::OR, MVT::i32, Tmp, OtherCond);
+ SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
+ return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
}
// Select - Convert the specified operand from a target-independent to a
// target-specific node if it hasn't already been changed.
SDNode *PPCDAGToDAGISel::Select(SDNode *N) {
- DebugLoc dl = N->getDebugLoc();
- if (N->isMachineOpcode())
+ SDLoc dl(N);
+ if (N->isMachineOpcode()) {
+ N->setNodeId(-1);
return NULL; // Already selected.
+ }
switch (N->getOpcode()) {
default: break;
// If it can't be represented as a 32 bit value.
if (!isInt<32>(Imm)) {
- Shift = CountTrailingZeros_64(Imm);
+ Shift = countTrailingZeros<uint64_t>(Imm);
int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift;
// If the shifted value fits 32 bits.
break;
}
- case ISD::SETCC:
- return SelectSETCC(N);
+ case ISD::SETCC: {
+ SDNode *SN = SelectSETCC(N);
+ if (SN)
+ return SN;
+ break;
+ }
case PPCISD::GlobalBaseReg:
return getGlobalBaseReg();
getSmallIPtrImm(0));
}
- case PPCISD::MFCR: {
+ case PPCISD::MFOCRF: {
SDValue InFlag = N->getOperand(1);
- // Use MFOCRF if supported.
- if (PPCSubTarget.hasMFOCRF())
- return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
- N->getOperand(0), InFlag);
- else
- return CurDAG->getMachineNode(PPC::MFCRpseud, dl, MVT::i32,
- N->getOperand(0), InFlag);
+ return CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32,
+ N->getOperand(0), InFlag);
}
case ISD::SDIV: {
break;
SDValue Offset = LD->getOffset();
- if (isa<ConstantSDNode>(Offset) ||
+ if (Offset.getOpcode() == ISD::TargetConstant ||
Offset.getOpcode() == ISD::TargetGlobalAddress) {
unsigned Opcode;
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);
+ MVT::Other, Ops);
} 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);
}
}
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
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);
+ SH = 0;
+
+ // If the operand is a logical right shift, we can fold it into this
+ // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb)
+ // for n <= mb. The right shift is really a left rotate followed by a
+ // mask, and this mask is a more-restrictive sub-mask of the mask implied
+ // by the shift.
+ if (Val.getOpcode() == ISD::SRL &&
+ isInt32Immediate(Val.getOperand(1).getNode(), Imm) && Imm <= MB) {
+ assert(Imm < 64 && "Illegal shift amount");
+ Val = Val.getOperand(0);
+ SH = 64 - Imm;
+ }
+
+ SDValue Ops[] = { Val, getI32Imm(SH), 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));
SDValue Ops[] = { N->getOperand(0).getOperand(0),
N->getOperand(0).getOperand(1),
getI32Imm(0), getI32Imm(MB),getI32Imm(ME) };
- return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops, 5);
+ return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
}
}
// Other cases are autogenerated.
break;
}
+ // FIXME: Remove this once the ANDI glue bug is fixed:
+ case PPCISD::ANDIo_1_EQ_BIT:
+ case PPCISD::ANDIo_1_GT_BIT: {
+ if (!ANDIGlueBug)
+ break;
+
+ EVT InVT = N->getOperand(0).getValueType();
+ assert((InVT == MVT::i64 || InVT == MVT::i32) &&
+ "Invalid input type for ANDIo_1_EQ_BIT");
+
+ unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDIo8 : PPC::ANDIo;
+ SDValue AndI(CurDAG->getMachineNode(Opcode, dl, InVT, MVT::Glue,
+ N->getOperand(0),
+ CurDAG->getTargetConstant(1, InVT)), 0);
+ SDValue CR0Reg = CurDAG->getRegister(PPC::CR0, MVT::i32);
+ SDValue SRIdxVal =
+ CurDAG->getTargetConstant(N->getOpcode() == PPCISD::ANDIo_1_EQ_BIT ?
+ PPC::sub_eq : PPC::sub_gt, MVT::i32);
+
+ return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1,
+ CR0Reg, SRIdxVal,
+ SDValue(AndI.getNode(), 1) /* glue */);
+ }
case ISD::SELECT_CC: {
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
bool isPPC64 = (PtrVT == MVT::i64);
+ // If this is a select of i1 operands, we'll pattern match it.
+ if (PPCSubTarget.useCRBits() &&
+ N->getOperand(0).getValueType() == MVT::i1)
+ break;
+
// Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc
if (!isPPC64)
if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
}
SDValue CCReg = SelectCC(N->getOperand(0), N->getOperand(1), CC, dl);
+
+ if (N->getValueType(0) == MVT::i1) {
+ // An i1 select is: (c & t) | (!c & f).
+ bool Inv;
+ unsigned Idx = getCRIdxForSetCC(CC, Inv);
+
+ unsigned SRI;
+ switch (Idx) {
+ default: llvm_unreachable("Invalid CC index");
+ case 0: SRI = PPC::sub_lt; break;
+ case 1: SRI = PPC::sub_gt; break;
+ case 2: SRI = PPC::sub_eq; break;
+ case 3: SRI = PPC::sub_un; break;
+ }
+
+ SDValue CCBit = CurDAG->getTargetExtractSubreg(SRI, dl, MVT::i1, CCReg);
+
+ SDValue NotCCBit(CurDAG->getMachineNode(PPC::CRNOR, dl, MVT::i1,
+ CCBit, CCBit), 0);
+ SDValue C = Inv ? NotCCBit : CCBit,
+ NotC = Inv ? CCBit : NotCCBit;
+
+ SDValue CAndT(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
+ C, N->getOperand(2)), 0);
+ SDValue NotCAndF(CurDAG->getMachineNode(PPC::CRAND, dl, MVT::i1,
+ NotC, N->getOperand(3)), 0);
+
+ return CurDAG->SelectNodeTo(N, PPC::CROR, MVT::i1, CAndT, NotCAndF);
+ }
+
unsigned BROpc = getPredicateForSetCC(CC);
unsigned SelectCCOp;
getI32Imm(BROpc) };
return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops, 4);
}
+ case PPCISD::BDNZ:
+ case PPCISD::BDZ: {
+ 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);
+ }
case PPCISD::COND_BRANCH: {
// Op #0 is the Chain.
// Op #1 is the PPC::PRED_* number.
}
case ISD::BR_CC: {
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(1))->get();
+ unsigned PCC = getPredicateForSetCC(CC);
+
+ if (N->getOperand(2).getValueType() == MVT::i1) {
+ unsigned Opc;
+ bool Swap;
+ switch (PCC) {
+ default: llvm_unreachable("Unexpected Boolean-operand predicate");
+ case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break;
+ case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break;
+ case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break;
+ case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break;
+ case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break;
+ case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break;
+ }
+
+ SDValue BitComp(CurDAG->getMachineNode(Opc, dl, MVT::i1,
+ N->getOperand(Swap ? 3 : 2),
+ N->getOperand(Swap ? 2 : 3)), 0);
+ return CurDAG->SelectNodeTo(N, PPC::BC, MVT::Other,
+ BitComp, N->getOperand(4), N->getOperand(0));
+ }
+
SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
- SDValue Ops[] = { getI32Imm(getPredicateForSetCC(CC)), CondCode,
+ SDValue Ops[] = { getI32Imm(PCC), CondCode,
N->getOperand(4), N->getOperand(0) };
return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops, 4);
}
Chain), 0);
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;
+
+ PeepholePPC64();
+ PeepholdCROps();
+}
+
+// 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::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::PeepholdCROps() {
+ bool IsModified;
+ do {
+ IsModified = false;
+ for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
+ E = CurDAG->allnodes_end(); I != E; ++I) {
+ MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(I);
+ if (!MachineNode || MachineNode->use_empty())
+ continue;
+ SDNode *ResNode = MachineNode;
+
+ bool Op1Set = false, Op1Unset = false,
+ Op1Not = false,
+ Op2Set = false, Op2Unset = false,
+ Op2Not = false;
+
+ unsigned Opcode = MachineNode->getMachineOpcode();
+ switch (Opcode) {
+ default: break;
+ case PPC::CRAND:
+ case PPC::CRNAND:
+ case PPC::CROR:
+ case PPC::CRXOR:
+ case PPC::CRNOR:
+ case PPC::CREQV:
+ case PPC::CRANDC:
+ case PPC::CRORC: {
+ SDValue Op = MachineNode->getOperand(1);
+ if (Op.isMachineOpcode()) {
+ if (Op.getMachineOpcode() == PPC::CRSET)
+ Op2Set = true;
+ else if (Op.getMachineOpcode() == PPC::CRUNSET)
+ Op2Unset = true;
+ else if (Op.getMachineOpcode() == PPC::CRNOR &&
+ Op.getOperand(0) == Op.getOperand(1))
+ Op2Not = true;
+ }
+ } // fallthrough
+ case PPC::BC:
+ case PPC::BCn:
+ case PPC::SELECT_I4:
+ case PPC::SELECT_I8:
+ case PPC::SELECT_F4:
+ case PPC::SELECT_F8:
+ case PPC::SELECT_VRRC: {
+ SDValue Op = MachineNode->getOperand(0);
+ if (Op.isMachineOpcode()) {
+ if (Op.getMachineOpcode() == PPC::CRSET)
+ Op1Set = true;
+ else if (Op.getMachineOpcode() == PPC::CRUNSET)
+ Op1Unset = true;
+ else if (Op.getMachineOpcode() == PPC::CRNOR &&
+ Op.getOperand(0) == Op.getOperand(1))
+ Op1Not = true;
+ }
+ }
+ break;
+ }
+
+ bool SelectSwap = false;
+ switch (Opcode) {
+ default: break;
+ case PPC::CRAND:
+ if (MachineNode->getOperand(0) == MachineNode->getOperand(1))
+ // x & x = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Set)
+ // 1 & y = y
+ ResNode = MachineNode->getOperand(1).getNode();
+ else if (Op2Set)
+ // x & 1 = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Unset || Op2Unset)
+ // x & 0 = 0 & y = 0
+ ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Not)
+ // ~x & y = andc(y, x)
+ ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(0).
+ getOperand(0));
+ else if (Op2Not)
+ // x & ~y = andc(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
+ 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))
+ // nand(x, x) -> nor(x, x)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(0));
+ else if (Op1Set)
+ // nand(1, y) -> nor(y, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(1));
+ else if (Op2Set)
+ // nand(x, 1) -> nor(x, x)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(0));
+ else if (Op1Unset || Op2Unset)
+ // nand(x, 0) = nand(0, y) = 1
+ ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Not)
+ // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(1));
+ else if (Op2Not)
+ // nand(x, ~y) = ~x | y = orc(y, x)
+ ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
+ 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))
+ // x | x = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Set || Op2Set)
+ // x | 1 = 1 | y = 1
+ ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Unset)
+ // 0 | y = y
+ ResNode = MachineNode->getOperand(1).getNode();
+ else if (Op2Unset)
+ // x | 0 = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Not)
+ // ~x | y = orc(y, x)
+ ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(0).
+ getOperand(0));
+ else if (Op2Not)
+ // x | ~y = orc(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRORC, SDLoc(MachineNode),
+ 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))
+ // xor(x, x) = 0
+ ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Set)
+ // xor(1, y) -> nor(y, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(1));
+ else if (Op2Set)
+ // xor(x, 1) -> nor(x, x)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(0));
+ else if (Op1Unset)
+ // xor(0, y) = y
+ ResNode = MachineNode->getOperand(1).getNode();
+ else if (Op2Unset)
+ // xor(x, 0) = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Not)
+ // xor(~x, y) = eqv(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(1));
+ else if (Op2Not)
+ // xor(x, ~y) = eqv(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CREQV, SDLoc(MachineNode),
+ 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)
+ // nor(1, y) -> 0
+ ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Unset)
+ // nor(0, y) = ~y -> nor(y, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(1));
+ else if (Op2Unset)
+ // nor(x, 0) = ~x
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(0));
+ else if (Op1Not)
+ // nor(~x, y) = andc(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(1));
+ else if (Op2Not)
+ // nor(x, ~y) = andc(y, x)
+ ResNode = CurDAG->getMachineNode(PPC::CRANDC, SDLoc(MachineNode),
+ 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))
+ // eqv(x, x) = 1
+ ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Set)
+ // eqv(1, y) = y
+ ResNode = MachineNode->getOperand(1).getNode();
+ else if (Op2Set)
+ // eqv(x, 1) = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Unset)
+ // eqv(0, y) = ~y -> nor(y, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(1));
+ else if (Op2Unset)
+ // eqv(x, 0) = ~x
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0),
+ MachineNode->getOperand(0));
+ else if (Op1Not)
+ // eqv(~x, y) = xor(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(1));
+ else if (Op2Not)
+ // eqv(x, ~y) = xor(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRXOR, SDLoc(MachineNode),
+ 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))
+ // andc(x, x) = 0
+ ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Set)
+ // andc(1, y) = ~y
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(1));
+ else if (Op1Unset || Op2Set)
+ // andc(0, y) = andc(x, 1) = 0
+ ResNode = CurDAG->getMachineNode(PPC::CRUNSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op2Unset)
+ // andc(x, 0) = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Not)
+ // andc(~x, y) = ~(x | y) = nor(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(1));
+ else if (Op2Not)
+ // andc(x, ~y) = x & y
+ ResNode = CurDAG->getMachineNode(PPC::CRAND, SDLoc(MachineNode),
+ 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))
+ // orc(x, x) = 1
+ ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op1Set || Op2Unset)
+ // orc(1, y) = orc(x, 0) = 1
+ ResNode = CurDAG->getMachineNode(PPC::CRSET, SDLoc(MachineNode),
+ MVT::i1);
+ else if (Op2Set)
+ // orc(x, 1) = x
+ ResNode = MachineNode->getOperand(0).getNode();
+ else if (Op1Unset)
+ // orc(0, y) = ~y
+ ResNode = CurDAG->getMachineNode(PPC::CRNOR, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(1),
+ MachineNode->getOperand(1));
+ else if (Op1Not)
+ // orc(~x, y) = ~(x & y) = nand(x, y)
+ ResNode = CurDAG->getMachineNode(PPC::CRNAND, SDLoc(MachineNode),
+ MVT::i1, MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(1));
+ else if (Op2Not)
+ // orc(x, ~y) = x | y
+ ResNode = CurDAG->getMachineNode(PPC::CROR, SDLoc(MachineNode),
+ 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:
+ if (Op1Set)
+ ResNode = MachineNode->getOperand(1).getNode();
+ else if (Op1Unset)
+ ResNode = MachineNode->getOperand(2).getNode();
+ else if (Op1Not)
+ ResNode = CurDAG->getMachineNode(MachineNode->getMachineOpcode(),
+ SDLoc(MachineNode),
+ MachineNode->getValueType(0),
+ MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(2),
+ MachineNode->getOperand(1));
+ break;
+ case PPC::BC:
+ case PPC::BCn:
+ if (Op1Not)
+ ResNode = CurDAG->getMachineNode(Opcode == PPC::BC ? PPC::BCn :
+ PPC::BC,
+ SDLoc(MachineNode),
+ MVT::Other,
+ MachineNode->getOperand(0).
+ getOperand(0),
+ MachineNode->getOperand(1),
+ MachineNode->getOperand(2));
+ // FIXME: Handle Op1Set, Op1Unset here too.
+ 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));
+ DEBUG(dbgs() << "\nNew: ");
+ DEBUG(ResNode->dump(CurDAG));
+ DEBUG(dbgs() << "\n");
+
+ ReplaceUses(MachineNode, ResNode);
+ IsModified = true;
+ }
+ }
+ if (IsModified)
+ CurDAG->RemoveDeadNodes();
+ } while (IsModified);
+}
+
+void PPCDAGToDAGISel::PeepholePPC64() {
+ // 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;
+
+ 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::ADDI:
+ // 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_DTPREL_LO;
+ break;
+ case PPC::ADDItlsldL:
+ Flags = PPCII::MO_TLSLD_LO;
+ break;
+ case PPC::ADDItocL:
+ Flags = PPCII::MO_TOC_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)) {
+ SDLoc dl(GA);
+ const GlobalValue *GV = GA->getGlobal();
+ // We can't perform this optimization for data whose alignment
+ // is insufficient for the instruction encoding.
+ if (GV->getAlignment() < 4 &&
+ (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
+ StorageOpcode == PPC::LWA)) {
+ DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
+ continue;
+ }
+ 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
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);
+}
+
projects / oota-llvm.git / blobdiff