cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug",
cl::desc("expose the ANDI glue bug on PPC"), cl::Hidden);
-cl::opt<bool> UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true),
- cl::desc("use aggressive ppc isel for bit permutations"), cl::Hidden);
-cl::opt<bool> BPermRewriterNoMasking("ppc-bit-perm-rewriter-stress-rotates",
- cl::desc("stress rotate selection in aggressive ppc isel for "
- "bit permutations"), cl::Hidden);
+static cl::opt<bool>
+ UseBitPermRewriter("ppc-use-bit-perm-rewriter", cl::init(true),
+ cl::desc("use aggressive ppc isel for bit permutations"),
+ cl::Hidden);
+static cl::opt<bool> BPermRewriterNoMasking(
+ "ppc-bit-perm-rewriter-stress-rotates",
+ cl::desc("stress rotate selection in aggressive ppc isel for "
+ "bit permutations"),
+ cl::Hidden);
namespace llvm {
void initializePPCDAGToDAGISelPass(PassRegistry&);
///
class PPCDAGToDAGISel : public SelectionDAGISel {
const PPCTargetMachine &TM;
- const PPCTargetLowering *PPCLowering;
const PPCSubtarget *PPCSubTarget;
+ const PPCTargetLowering *PPCLowering;
unsigned GlobalBaseReg;
public:
explicit PPCDAGToDAGISel(PPCTargetMachine &tm)
- : SelectionDAGISel(tm), TM(tm),
- PPCLowering(TM.getSubtargetImpl()->getTargetLowering()),
- PPCSubTarget(TM.getSubtargetImpl()) {
+ : SelectionDAGISel(tm), TM(tm) {
initializePPCDAGToDAGISelPass(*PassRegistry::getPassRegistry());
}
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();
+ PPCSubTarget = &MF.getSubtarget<PPCSubtarget>();
+ PPCLowering = PPCSubTarget->getTargetLowering();
SelectionDAGISel::runOnMachineFunction(MF);
if (!PPCSubTarget->isSVR4ABI())
return true;
}
+ void PreprocessISelDAG() override;
void PostprocessISelDAG() override;
/// getI32Imm - Return a target constant with the specified value, of type
/// i32.
- inline SDValue getI32Imm(unsigned Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i32);
+ inline SDValue getI32Imm(unsigned Imm, SDLoc dl) {
+ return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
}
/// getI64Imm - Return a target constant with the specified value, of type
/// i64.
- inline SDValue getI64Imm(uint64_t Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i64);
+ inline SDValue getI64Imm(uint64_t Imm, SDLoc dl) {
+ return CurDAG->getTargetConstant(Imm, dl, MVT::i64);
}
/// getSmallIPtrImm - Return a target constant of pointer type.
- inline SDValue getSmallIPtrImm(unsigned Imm) {
- return CurDAG->getTargetConstant(Imm, PPCLowering->getPointerTy());
+ inline SDValue getSmallIPtrImm(unsigned Imm, SDLoc dl) {
+ return CurDAG->getTargetConstant(
+ Imm, dl, PPCLowering->getPointerTy(CurDAG->getDataLayout()));
}
- /// isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s
- /// with any number of 0s on either side. The 1s are allowed to wrap from
- /// LSB to MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs.
- /// 0x0F0F0000 is not, since all 1s are not contiguous.
- static bool isRunOfOnes(unsigned Val, unsigned &MB, unsigned &ME);
-
-
/// isRotateAndMask - Returns true if Mask and Shift can be folded into a
/// rotate and mask opcode and mask operation.
static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask,
/// 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.
bool SelectInlineAsmMemoryOperand(const SDValue &Op,
- char ConstraintCode,
+ unsigned ConstraintID,
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;
+
+ switch(ConstraintID) {
+ default:
+ errs() << "ConstraintID: " << ConstraintID << "\n";
+ llvm_unreachable("Unexpected asm memory constraint");
+ case InlineAsm::Constraint_es:
+ case InlineAsm::Constraint_i:
+ case InlineAsm::Constraint_m:
+ case InlineAsm::Constraint_o:
+ case InlineAsm::Constraint_Q:
+ case InlineAsm::Constraint_Z:
+ case InlineAsm::Constraint_Zy:
+ // 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 = PPCSubTarget->getRegisterInfo();
+ const TargetRegisterClass *TRC = TRI->getPointerRegClass(*MF, /*Kind=*/1);
+ SDLoc dl(Op);
+ SDValue RC = CurDAG->getTargetConstant(TRC->getID(), dl, MVT::i32);
+ SDValue NewOp =
+ SDValue(CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
+ dl, Op.getValueType(),
+ Op, RC), 0);
+
+ OutOps.push_back(NewOp);
+ return false;
+ }
+ return true;
}
void InsertVRSaveCode(MachineFunction &MF);
void PeepholePPC64ZExt();
void PeepholeCROps();
+ SDValue combineToCMPB(SDNode *N);
+ void foldBoolExts(SDValue &Res, SDNode *&N);
+
bool AllUsersSelectZero(SDNode *N);
void SwapAllSelectUsers(SDNode *N);
+
+ SDNode *transferMemOperands(SDNode *N, SDNode *Result);
};
}
unsigned InVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
unsigned UpdatedVRSAVE = RegInfo->createVirtualRegister(&PPC::GPRCRegClass);
- const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
+ const TargetInstrInfo &TII = *PPCSubTarget->getInstrInfo();
MachineBasicBlock &EntryBB = *Fn.begin();
DebugLoc dl;
// Emit the following code into the entry block:
// Find all return blocks, outputting a restore in each epilog.
for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
- if (!BB->empty() && BB->back().isReturn()) {
+ if (BB->isReturnBlock()) {
IP = BB->end(); --IP;
// Skip over all terminator instructions, which are part of the return
///
SDNode *PPCDAGToDAGISel::getGlobalBaseReg() {
if (!GlobalBaseReg) {
- const TargetInstrInfo &TII = *TM.getSubtargetImpl()->getInstrInfo();
+ const TargetInstrInfo &TII = *PPCSubTarget->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) {
+ if (PPCLowering->getPointerTy(CurDAG->getDataLayout()) == 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);
+ MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
} 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)
+ TII.get(PPC::UpdateGBR), GlobalBaseReg)
.addReg(TempReg, RegState::Define).addReg(GlobalBaseReg);
MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true);
}
}
}
return CurDAG->getRegister(GlobalBaseReg,
- PPCLowering->getPointerTy()).getNode();
+ PPCLowering->getPointerTy(CurDAG->getDataLayout()))
+ .getNode();
}
/// isIntS16Immediate - This method tests to see if the node is either a 32-bit
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));
+ getSmallIPtrImm(Offset, dl));
return CurDAG->getMachineNode(Opc, dl, N->getValueType(0), TFI,
- getSmallIPtrImm(Offset));
-}
-
-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(Val);
- // look for the first zero bit after the run of ones
- 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(Val) - 1;
- // effectively look for the first one bit after the run of zeros
- MB = countLeadingZeros((Val - 1) ^ Val) + 1;
- return true;
- }
- }
- // no run present
- return false;
+ getSmallIPtrImm(Offset, dl));
}
bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask,
}
SH &= 31;
- SDValue Ops[] = { Op0, Op1, getI32Imm(SH), getI32Imm(MB),
- getI32Imm(ME) };
+ SDValue Ops[] = { Op0, Op1, getI32Imm(SH, dl), getI32Imm(MB, dl),
+ getI32Imm(ME, dl) };
return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
}
}
// Predict the number of instructions that would be generated by calling
// SelectInt64(N).
-static unsigned SelectInt64Count(int64_t Imm) {
+static unsigned SelectInt64CountDirect(int64_t Imm) {
// Assume no remaining bits.
unsigned Remainder = 0;
// Assume no shift required.
// Handle first 32 bits.
unsigned Lo = Imm & 0xFFFF;
- unsigned Hi = (Imm >> 16) & 0xFFFF;
// Simple value.
if (isInt<16>(Imm)) {
++Result;
// Add in the last bits as required.
- if ((Hi = (Remainder >> 16) & 0xFFFF))
+ if ((Remainder >> 16) & 0xFFFF)
++Result;
- if ((Lo = Remainder & 0xFFFF))
+ if (Remainder & 0xFFFF)
++Result;
return Result;
}
+static uint64_t Rot64(uint64_t Imm, unsigned R) {
+ return (Imm << R) | (Imm >> (64 - R));
+}
+
+static unsigned SelectInt64Count(int64_t Imm) {
+ unsigned Count = SelectInt64CountDirect(Imm);
+ if (Count == 1)
+ return Count;
+
+ for (unsigned r = 1; r < 63; ++r) {
+ uint64_t RImm = Rot64(Imm, r);
+ unsigned RCount = SelectInt64CountDirect(RImm) + 1;
+ Count = std::min(Count, RCount);
+
+ // See comments in SelectInt64 for an explanation of the logic below.
+ unsigned LS = findLastSet(RImm);
+ if (LS != r-1)
+ continue;
+
+ uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1));
+ uint64_t RImmWithOnes = RImm | OnesMask;
+
+ RCount = SelectInt64CountDirect(RImmWithOnes) + 1;
+ Count = std::min(Count, RCount);
+ }
+
+ return Count;
+}
+
// Select a 64-bit constant. For cost-modeling purposes, SelectInt64Count
// (above) needs to be kept in sync with this function.
-static SDNode *SelectInt64(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) {
+static SDNode *SelectInt64Direct(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) {
// Assume no remaining bits.
unsigned Remainder = 0;
// Assume no shift required.
unsigned Lo = Imm & 0xFFFF;
unsigned Hi = (Imm >> 16) & 0xFFFF;
- auto getI32Imm = [CurDAG](unsigned Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i32);
+ auto getI32Imm = [CurDAG, dl](unsigned Imm) {
+ return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
};
// Simple value.
return Result;
}
+static SDNode *SelectInt64(SelectionDAG *CurDAG, SDLoc dl, int64_t Imm) {
+ unsigned Count = SelectInt64CountDirect(Imm);
+ if (Count == 1)
+ return SelectInt64Direct(CurDAG, dl, Imm);
+
+ unsigned RMin = 0;
+
+ int64_t MatImm;
+ unsigned MaskEnd;
+
+ for (unsigned r = 1; r < 63; ++r) {
+ uint64_t RImm = Rot64(Imm, r);
+ unsigned RCount = SelectInt64CountDirect(RImm) + 1;
+ if (RCount < Count) {
+ Count = RCount;
+ RMin = r;
+ MatImm = RImm;
+ MaskEnd = 63;
+ }
+
+ // If the immediate to generate has many trailing zeros, it might be
+ // worthwhile to generate a rotated value with too many leading ones
+ // (because that's free with li/lis's sign-extension semantics), and then
+ // mask them off after rotation.
+
+ unsigned LS = findLastSet(RImm);
+ // We're adding (63-LS) higher-order ones, and we expect to mask them off
+ // after performing the inverse rotation by (64-r). So we need that:
+ // 63-LS == 64-r => LS == r-1
+ if (LS != r-1)
+ continue;
+
+ uint64_t OnesMask = -(int64_t) (UINT64_C(1) << (LS+1));
+ uint64_t RImmWithOnes = RImm | OnesMask;
+
+ RCount = SelectInt64CountDirect(RImmWithOnes) + 1;
+ if (RCount < Count) {
+ Count = RCount;
+ RMin = r;
+ MatImm = RImmWithOnes;
+ MaskEnd = LS;
+ }
+ }
+
+ if (!RMin)
+ return SelectInt64Direct(CurDAG, dl, Imm);
+
+ auto getI32Imm = [CurDAG, dl](unsigned Imm) {
+ return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
+ };
+
+ SDValue Val = SDValue(SelectInt64Direct(CurDAG, dl, MatImm), 0);
+ return CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Val,
+ getI32Imm(64 - RMin), getI32Imm(MaskEnd));
+}
+
// Select a 64-bit constant.
static SDNode *SelectInt64(SelectionDAG *CurDAG, SDNode *N) {
SDLoc dl(N);
return SelectInt64(CurDAG, dl, Imm);
}
-
namespace {
class BitPermutationSelector {
struct ValueBit {
BitGroups[BitGroups.size()-1].EndIdx == Bits.size()-1 &&
BitGroups[0].V == BitGroups[BitGroups.size()-1].V &&
BitGroups[0].RLAmt == BitGroups[BitGroups.size()-1].RLAmt) {
- DEBUG(dbgs() << "\tcombining final bit group with inital one\n");
+ DEBUG(dbgs() << "\tcombining final bit group with initial one\n");
BitGroups[BitGroups.size()-1].EndIdx = BitGroups[0].EndIdx;
BitGroups.erase(BitGroups.begin());
}
}
}
- SDValue getI32Imm(unsigned Imm) {
- return CurDAG->getTargetConstant(Imm, MVT::i32);
+ SDValue getI32Imm(unsigned Imm, SDLoc dl) {
+ return CurDAG->getTargetConstant(Imm, dl, MVT::i32);
}
uint64_t getZerosMask() {
for (unsigned i = 0; i < Bits.size(); ++i) {
if (Bits[i].hasValue())
continue;
- Mask |= (1ul << i);
+ Mask |= (UINT64_C(1) << i);
}
return ~Mask;
SDValue VRot;
if (VRI.RLAmt) {
SDValue Ops[] =
- { VRI.V, getI32Imm(VRI.RLAmt), getI32Imm(0), getI32Imm(31) };
+ { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl),
+ getI32Imm(31, dl) };
VRot = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32,
Ops), 0);
} else {
SDValue ANDIVal, ANDISVal;
if (ANDIMask != 0)
ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
- VRot, getI32Imm(ANDIMask)), 0);
+ VRot, getI32Imm(ANDIMask, dl)), 0);
if (ANDISMask != 0)
ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
- VRot, getI32Imm(ANDISMask)), 0);
+ VRot, getI32Imm(ANDISMask, dl)), 0);
SDValue TotalVal;
if (!ANDIVal)
// Now, remove all groups with this underlying value and rotation
// factor.
- for (auto I = BitGroups.begin(); I != BitGroups.end();) {
- if (I->V == VRI.V && I->RLAmt == VRI.RLAmt)
- I = BitGroups.erase(I);
- else
- ++I;
- }
+ eraseMatchingBitGroups([VRI](const BitGroup &BG) {
+ return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt;
+ });
}
}
if (VRI.RLAmt) {
if (InstCnt) *InstCnt += 1;
SDValue Ops[] =
- { VRI.V, getI32Imm(VRI.RLAmt), getI32Imm(0), getI32Imm(31) };
- Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
+ { VRI.V, getI32Imm(VRI.RLAmt, dl), getI32Imm(0, dl),
+ getI32Imm(31, dl) };
+ Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops),
+ 0);
} else {
Res = VRI.V;
}
// Now, remove all groups with this underlying value and rotation factor.
- for (auto I = BitGroups.begin(); I != BitGroups.end();) {
- if (I->V == VRI.V && I->RLAmt == VRI.RLAmt)
- I = BitGroups.erase(I);
- else
- ++I;
- }
+ eraseMatchingBitGroups([VRI](const BitGroup &BG) {
+ return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt;
+ });
}
if (InstCnt) *InstCnt += BitGroups.size();
for (auto &BG : BitGroups) {
if (!Res) {
SDValue Ops[] =
- { BG.V, getI32Imm(BG.RLAmt), getI32Imm(Bits.size() - BG.EndIdx - 1),
- getI32Imm(Bits.size() - BG.StartIdx - 1) };
+ { BG.V, getI32Imm(BG.RLAmt, dl),
+ getI32Imm(Bits.size() - BG.EndIdx - 1, dl),
+ getI32Imm(Bits.size() - BG.StartIdx - 1, dl) };
Res = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
} else {
SDValue Ops[] =
- { Res, BG.V, getI32Imm(BG.RLAmt), getI32Imm(Bits.size() - BG.EndIdx - 1),
- getI32Imm(Bits.size() - BG.StartIdx - 1) };
+ { Res, BG.V, getI32Imm(BG.RLAmt, dl),
+ getI32Imm(Bits.size() - BG.EndIdx - 1, dl),
+ getI32Imm(Bits.size() - BG.StartIdx - 1, dl) };
Res = SDValue(CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops), 0);
}
}
SDValue ANDIVal, ANDISVal;
if (ANDIMask != 0)
ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo, dl, MVT::i32,
- Res, getI32Imm(ANDIMask)), 0);
+ Res, getI32Imm(ANDIMask, dl)), 0);
if (ANDISMask != 0)
ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo, dl, MVT::i32,
- Res, getI32Imm(ANDISMask)), 0);
+ Res, getI32Imm(ANDISMask, dl)), 0);
if (!ANDIVal)
Res = ANDISVal;
assert(InstMaskStart >= 32 && "Mask cannot start out of range");
assert(InstMaskEnd >= 32 && "Mask cannot end out of range");
SDValue Ops[] =
- { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart - 32),
- getI32Imm(InstMaskEnd - 32) };
+ { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl),
+ getI32Imm(InstMaskEnd - 32, dl) };
return SDValue(CurDAG->getMachineNode(PPC::RLWINM8, dl, MVT::i64,
Ops), 0);
}
if (InstMaskEnd == 63) {
SDValue Ops[] =
- { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) };
+ { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
return SDValue(CurDAG->getMachineNode(PPC::RLDICL, dl, MVT::i64, Ops), 0);
}
if (InstMaskStart == 0) {
SDValue Ops[] =
- { V, getI32Imm(RLAmt), getI32Imm(InstMaskEnd) };
+ { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskEnd, dl) };
return SDValue(CurDAG->getMachineNode(PPC::RLDICR, dl, MVT::i64, Ops), 0);
}
if (InstMaskEnd == 63 - RLAmt) {
SDValue Ops[] =
- { V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) };
+ { V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
return SDValue(CurDAG->getMachineNode(PPC::RLDIC, dl, MVT::i64, Ops), 0);
}
assert(InstMaskStart >= 32 && "Mask cannot start out of range");
assert(InstMaskEnd >= 32 && "Mask cannot end out of range");
SDValue Ops[] =
- { Base, V, getI32Imm(RLAmt), getI32Imm(InstMaskStart - 32),
- getI32Imm(InstMaskEnd - 32) };
+ { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart - 32, dl),
+ getI32Imm(InstMaskEnd - 32, dl) };
return SDValue(CurDAG->getMachineNode(PPC::RLWIMI8, dl, MVT::i64,
Ops), 0);
}
if (InstMaskEnd == 63 - RLAmt) {
SDValue Ops[] =
- { Base, V, getI32Imm(RLAmt), getI32Imm(InstMaskStart) };
+ { Base, V, getI32Imm(RLAmt, dl), getI32Imm(InstMaskStart, dl) };
return SDValue(CurDAG->getMachineNode(PPC::RLDIMI, dl, MVT::i64, Ops), 0);
}
// Repl32 true, but are trivially convertable to Repl32 false. Such a
// group is trivially convertable if it overlaps only with the lower 32
// bits, and the group has not been coalesced.
- auto MatchingBG = [VRI](BitGroup &BG) {
+ auto MatchingBG = [VRI](const BitGroup &BG) {
if (VRI.V != BG.V)
return false;
if (BG.StartIdx <= BG.EndIdx) {
for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i)
- Mask |= (1ul << i);
+ Mask |= (UINT64_C(1) << i);
} else {
for (unsigned i = BG.StartIdx; i < Bits.size(); ++i)
- Mask |= (1ul << i);
+ Mask |= (UINT64_C(1) << i);
for (unsigned i = 0; i <= BG.EndIdx; ++i)
- Mask |= (1ul << i);
+ Mask |= (UINT64_C(1) << i);
}
}
SDValue ANDIVal, ANDISVal;
if (ANDIMask != 0)
ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
- VRot, getI32Imm(ANDIMask)), 0);
+ VRot, getI32Imm(ANDIMask, dl)), 0);
if (ANDISMask != 0)
ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
- VRot, getI32Imm(ANDISMask)), 0);
+ VRot, getI32Imm(ANDISMask, dl)), 0);
if (!ANDIVal)
TotalVal = ANDISVal;
// Now, remove all groups with this underlying value and rotation
// factor.
- for (auto I = BitGroups.begin(); I != BitGroups.end();) {
- if (MatchingBG(*I))
- I = BitGroups.erase(I);
- else
- ++I;
- }
+ eraseMatchingBitGroups(MatchingBG);
}
}
// Now, remove all groups with this underlying value and rotation factor.
if (Res)
- for (auto I = BitGroups.begin(); I != BitGroups.end();) {
- if (I->V == VRI.V && I->RLAmt == VRI.RLAmt && I->Repl32 == VRI.Repl32)
- I = BitGroups.erase(I);
- else
- ++I;
- }
+ eraseMatchingBitGroups([VRI](const BitGroup &BG) {
+ return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt &&
+ BG.Repl32 == VRI.Repl32;
+ });
}
// Because 64-bit rotates are more flexible than inserts, we might have a
SDValue ANDIVal, ANDISVal;
if (ANDIMask != 0)
ANDIVal = SDValue(CurDAG->getMachineNode(PPC::ANDIo8, dl, MVT::i64,
- Res, getI32Imm(ANDIMask)), 0);
+ Res, getI32Imm(ANDIMask, dl)), 0);
if (ANDISMask != 0)
ANDISVal = SDValue(CurDAG->getMachineNode(PPC::ANDISo8, dl, MVT::i64,
- Res, getI32Imm(ANDISMask)), 0);
+ Res, getI32Imm(ANDISMask, dl)), 0);
if (!ANDIVal)
Res = ANDISVal;
return nullptr;
}
+ void eraseMatchingBitGroups(function_ref<bool(const BitGroup &)> F) {
+ BitGroups.erase(std::remove_if(BitGroups.begin(), BitGroups.end(), F),
+ BitGroups.end());
+ }
+
SmallVector<ValueBit, 64> Bits;
bool HasZeros;
// SETEQ/SETNE comparison with 16-bit immediate, fold it.
if (isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
- getI32Imm(Imm & 0xFFFF)), 0);
+ getI32Imm(Imm & 0xFFFF, dl)),
+ 0);
// If this is a 16-bit signed immediate, fold it.
if (isInt<16>((int)Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
- getI32Imm(Imm & 0xFFFF)), 0);
+ getI32Imm(Imm & 0xFFFF, dl)),
+ 0);
// For non-equality comparisons, the default code would materialize the
// constant, then compare against it, like this:
// cmplwi cr0,r0,0x5678
// beq cr0,L6
SDValue Xor(CurDAG->getMachineNode(PPC::XORIS, dl, MVT::i32, LHS,
- getI32Imm(Imm >> 16)), 0);
+ getI32Imm(Imm >> 16, dl)), 0);
return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, Xor,
- getI32Imm(Imm & 0xFFFF)), 0);
+ getI32Imm(Imm & 0xFFFF, dl)), 0);
}
Opc = PPC::CMPLW;
} else if (ISD::isUnsignedIntSetCC(CC)) {
if (isInt32Immediate(RHS, Imm) && isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLWI, dl, MVT::i32, LHS,
- getI32Imm(Imm & 0xFFFF)), 0);
+ getI32Imm(Imm & 0xFFFF, dl)), 0);
Opc = PPC::CMPLW;
} else {
short SImm;
if (isIntS16Immediate(RHS, SImm))
return SDValue(CurDAG->getMachineNode(PPC::CMPWI, dl, MVT::i32, LHS,
- getI32Imm((int)SImm & 0xFFFF)),
+ getI32Imm((int)SImm & 0xFFFF,
+ dl)),
0);
Opc = PPC::CMPW;
}
// SETEQ/SETNE comparison with 16-bit immediate, fold it.
if (isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
- getI32Imm(Imm & 0xFFFF)), 0);
+ getI32Imm(Imm & 0xFFFF, dl)),
+ 0);
// If this is a 16-bit signed immediate, fold it.
if (isInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
- getI32Imm(Imm & 0xFFFF)), 0);
+ getI32Imm(Imm & 0xFFFF, dl)),
+ 0);
// For non-equality comparisons, the default code would materialize the
// constant, then compare against it, like this:
// beq cr0,L6
if (isUInt<32>(Imm)) {
SDValue Xor(CurDAG->getMachineNode(PPC::XORIS8, dl, MVT::i64, LHS,
- getI64Imm(Imm >> 16)), 0);
+ getI64Imm(Imm >> 16, dl)), 0);
return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, Xor,
- getI64Imm(Imm & 0xFFFF)), 0);
+ getI64Imm(Imm & 0xFFFF, dl)),
+ 0);
}
}
Opc = PPC::CMPLD;
} else if (ISD::isUnsignedIntSetCC(CC)) {
if (isInt64Immediate(RHS.getNode(), Imm) && isUInt<16>(Imm))
return SDValue(CurDAG->getMachineNode(PPC::CMPLDI, dl, MVT::i64, LHS,
- getI64Imm(Imm & 0xFFFF)), 0);
+ getI64Imm(Imm & 0xFFFF, dl)), 0);
Opc = PPC::CMPLD;
} else {
short SImm;
if (isIntS16Immediate(RHS, SImm))
return SDValue(CurDAG->getMachineNode(PPC::CMPDI, dl, MVT::i64, LHS,
- getI64Imm(SImm & 0xFFFF)),
+ getI64Imm(SImm & 0xFFFF, dl)),
0);
Opc = PPC::CMPD;
}
// 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).
+// only support the altivec types (v16i8, v8i16, v4i32, v2i64, and v4f32).
static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC,
bool HasVSX, bool &Swap, bool &Negate) {
Swap = false;
return PPC::VCMPEQUH;
else if (VecVT == MVT::v4i32)
return PPC::VCMPEQUW;
+ else if (VecVT == MVT::v2i64)
+ return PPC::VCMPEQUD;
break;
case ISD::SETGT:
if (VecVT == MVT::v16i8)
return PPC::VCMPGTSH;
else if (VecVT == MVT::v4i32)
return PPC::VCMPGTSW;
+ else if (VecVT == MVT::v2i64)
+ return PPC::VCMPGTSD;
break;
case ISD::SETUGT:
if (VecVT == MVT::v16i8)
return PPC::VCMPGTUH;
else if (VecVT == MVT::v4i32)
return PPC::VCMPGTUW;
+ else if (VecVT == MVT::v2i64)
+ return PPC::VCMPGTUD;
break;
default:
break;
SDLoc dl(N);
unsigned Imm;
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
- EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
+ EVT PtrVT =
+ CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout());
bool isPPC64 = (PtrVT == MVT::i64);
if (!PPCSubTarget->useCRBits() &&
default: break;
case ISD::SETEQ: {
Op = SDValue(CurDAG->getMachineNode(PPC::CNTLZW, dl, MVT::i32, Op), 0);
- SDValue Ops[] = { Op, getI32Imm(27), getI32Imm(5), getI32Imm(31) };
+ SDValue Ops[] = { Op, getI32Imm(27, dl), getI32Imm(5, dl),
+ getI32Imm(31, dl) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
case ISD::SETNE: {
if (isPPC64) break;
SDValue AD =
SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
- Op, getI32Imm(~0U)), 0);
+ Op, getI32Imm(~0U, dl)), 0);
return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, AD, Op,
AD.getValue(1));
}
case ISD::SETLT: {
- SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
+ SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl),
+ getI32Imm(31, dl) };
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) };
+ SDValue Ops[] = { T, getI32Imm(1, dl), getI32Imm(31, dl),
+ getI32Imm(31, dl) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
}
case ISD::SETEQ:
if (isPPC64) break;
Op = SDValue(CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
- Op, getI32Imm(1)), 0);
+ Op, getI32Imm(1, dl)), 0);
return CurDAG->SelectNodeTo(N, PPC::ADDZE, MVT::i32,
SDValue(CurDAG->getMachineNode(PPC::LI, dl,
MVT::i32,
- getI32Imm(0)), 0),
- Op.getValue(1));
+ getI32Imm(0, dl)),
+ 0), Op.getValue(1));
case ISD::SETNE: {
if (isPPC64) break;
Op = SDValue(CurDAG->getMachineNode(PPC::NOR, dl, MVT::i32, Op, Op), 0);
SDNode *AD = CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
- Op, getI32Imm(~0U));
+ Op, getI32Imm(~0U, dl));
return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32, SDValue(AD, 0),
Op, SDValue(AD, 1));
}
case ISD::SETLT: {
SDValue AD = SDValue(CurDAG->getMachineNode(PPC::ADDI, dl, MVT::i32, Op,
- getI32Imm(1)), 0);
+ getI32Imm(1, dl)), 0);
SDValue AN = SDValue(CurDAG->getMachineNode(PPC::AND, dl, MVT::i32, AD,
Op), 0);
- SDValue Ops[] = { AN, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
+ SDValue Ops[] = { AN, getI32Imm(1, dl), getI32Imm(31, dl),
+ getI32Imm(31, dl) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
case ISD::SETGT: {
- SDValue Ops[] = { Op, getI32Imm(1), getI32Imm(31), getI32Imm(31) };
- Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops),
- 0);
+ SDValue Ops[] = { Op, getI32Imm(1, dl), getI32Imm(31, dl),
+ getI32Imm(31, dl) };
+ Op = SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Op,
- getI32Imm(1));
+ getI32Imm(1, dl));
}
}
}
// 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()) {
+ if (PPCSubTarget->hasQPX())
+ return nullptr;
+
EVT VecVT = LHS.getValueType();
bool Swap, Negate;
unsigned int VCmpInst = getVCmpInst(VecVT.getSimpleVT(), CC,
if (Swap)
std::swap(LHS, RHS);
+ EVT ResVT = VecVT.changeVectorElementTypeToInteger();
if (Negate) {
- SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, VecVT, LHS, RHS), 0);
+ SDValue VCmp(CurDAG->getMachineNode(VCmpInst, dl, ResVT, LHS, RHS), 0);
return CurDAG->SelectNodeTo(N, PPCSubTarget->hasVSX() ? PPC::XXLNOR :
PPC::VNOR,
- VecVT, VCmp, VCmp);
+ ResVT, VCmp, VCmp);
}
- return CurDAG->SelectNodeTo(N, VCmpInst, VecVT, LHS, RHS);
+ return CurDAG->SelectNodeTo(N, VCmpInst, ResVT, LHS, RHS);
}
if (PPCSubTarget->useCRBits())
IntCR = SDValue(CurDAG->getMachineNode(PPC::MFOCRF, dl, MVT::i32, CR7Reg,
CCReg), 0);
- SDValue Ops[] = { IntCR, getI32Imm((32-(3-Idx)) & 31),
- getI32Imm(31), getI32Imm(31) };
+ SDValue Ops[] = { IntCR, getI32Imm((32 - (3 - Idx)) & 31, dl),
+ getI32Imm(31, dl), getI32Imm(31, dl) };
if (!Inv)
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
// Get the specified bit.
SDValue Tmp =
SDValue(CurDAG->getMachineNode(PPC::RLWINM, dl, MVT::i32, Ops), 0);
- return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1));
+ return CurDAG->SelectNodeTo(N, PPC::XORI, MVT::i32, Tmp, getI32Imm(1, dl));
+}
+
+SDNode *PPCDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) {
+ // Transfer memoperands.
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
+ cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
+ return Result;
}
SDValue N0 = N->getOperand(0);
SDValue ShiftAmt =
CurDAG->getTargetConstant(*cast<ConstantSDNode>(N->getOperand(1))->
- getConstantIntValue(), N->getValueType(0));
+ getConstantIntValue(), dl,
+ N->getValueType(0));
if (N->getValueType(0) == MVT::i64) {
SDNode *Op =
CurDAG->getMachineNode(PPC::SRADI, dl, MVT::i64, MVT::Glue,
SDValue Chain = LD->getChain();
SDValue Base = LD->getBasePtr();
SDValue Ops[] = { Offset, Base, Chain };
- return CurDAG->getMachineNode(Opcode, dl, LD->getValueType(0),
- PPCLowering->getPointerTy(),
- MVT::Other, Ops);
+ return transferMemOperands(
+ N, CurDAG->getMachineNode(
+ Opcode, dl, LD->getValueType(0),
+ PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other,
+ Ops));
} else {
unsigned Opcode;
bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD;
assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load");
switch (LoadedVT.getSimpleVT().SimpleTy) {
default: llvm_unreachable("Invalid PPC load type!");
+ case MVT::v4f64: Opcode = PPC::QVLFDUX; break; // QPX
+ case MVT::v4f32: Opcode = PPC::QVLFSUX; break; // QPX
case MVT::f64: Opcode = PPC::LFDUX; break;
case MVT::f32: Opcode = PPC::LFSUX; break;
case MVT::i32: Opcode = PPC::LWZUX; 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);
+ return transferMemOperands(
+ N, CurDAG->getMachineNode(
+ Opcode, dl, LD->getValueType(0),
+ PPCLowering->getPointerTy(CurDAG->getDataLayout()), MVT::Other,
+ Ops));
}
}
if (isInt32Immediate(N->getOperand(1), Imm) &&
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) };
+ SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl),
+ getI32Imm(ME, dl) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
// If this is just a masked value where the input is not handled above, and
isRunOfOnes(Imm, MB, ME) &&
N->getOperand(0).getOpcode() != ISD::ROTL) {
SDValue Val = N->getOperand(0);
- SDValue Ops[] = { Val, getI32Imm(0), getI32Imm(MB), getI32Imm(ME) };
+ SDValue Ops[] = { Val, getI32Imm(0, dl), getI32Imm(MB, dl),
+ getI32Imm(ME, dl) };
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) &&
isMask_64(Imm64)) {
SDValue Val = N->getOperand(0);
- MB = 64 - CountTrailingOnes_64(Imm64);
+ MB = 64 - countTrailingOnes(Imm64);
SH = 0;
// If the operand is a logical right shift, we can fold it into this
SH = 64 - Imm;
}
- SDValue Ops[] = { Val, getI32Imm(SH), getI32Imm(MB) };
+ SDValue Ops[] = { Val, getI32Imm(SH, dl), getI32Imm(MB, dl) };
return CurDAG->SelectNodeTo(N, PPC::RLDICL, MVT::i64, Ops);
}
// AND X, 0 -> 0, not "rlwinm 32".
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
+ // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) might be a
+ // bitfield insert.
if (isInt32Immediate(N->getOperand(1), Imm) &&
N->getOperand(0).getOpcode() == ISD::OR &&
isInt32Immediate(N->getOperand(0).getOperand(1), Imm2)) {
+ // The idea here is to check whether this is equivalent to:
+ // (c1 & m) | (x & ~m)
+ // where m is a run-of-ones mask. The logic here is that, for each bit in
+ // c1 and c2:
+ // - if both are 1, then the output will be 1.
+ // - if both are 0, then the output will be 0.
+ // - if the bit in c1 is 0, and the bit in c2 is 1, then the output will
+ // come from x.
+ // - if the bit in c1 is 1, and the bit in c2 is 0, then the output will
+ // be 0.
+ // If that last condition is never the case, then we can form m from the
+ // bits that are the same between c1 and c2.
unsigned MB, ME;
- Imm = ~(Imm^Imm2);
- if (isRunOfOnes(Imm, MB, ME)) {
+ if (isRunOfOnes(~(Imm^Imm2), MB, ME) && !(~Imm & Imm2)) {
SDValue Ops[] = { N->getOperand(0).getOperand(0),
N->getOperand(0).getOperand(1),
- getI32Imm(0), getI32Imm(MB),getI32Imm(ME) };
+ getI32Imm(0, dl), getI32Imm(MB, dl),
+ getI32Imm(ME, dl) };
return CurDAG->getMachineNode(PPC::RLWIMI, dl, MVT::i32, Ops);
}
}
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) };
+ getI32Imm(SH, dl), getI32Imm(MB, dl),
+ getI32Imm(ME, dl) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
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) };
+ getI32Imm(SH, dl), getI32Imm(MB, dl),
+ getI32Imm(ME, dl) };
return CurDAG->SelectNodeTo(N, PPC::RLWINM, MVT::i32, Ops);
}
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);
+ CurDAG->getTargetConstant(1, dl, 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);
+ PPC::sub_eq : PPC::sub_gt, dl, MVT::i32);
return CurDAG->SelectNodeTo(N, TargetOpcode::EXTRACT_SUBREG, MVT::i1,
CR0Reg, SRIdxVal,
}
case ISD::SELECT_CC: {
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(4))->get();
- EVT PtrVT = CurDAG->getTargetLoweringInfo().getPointerTy();
+ EVT PtrVT =
+ CurDAG->getTargetLoweringInfo().getPointerTy(CurDAG->getDataLayout());
bool isPPC64 = (PtrVT == MVT::i64);
// If this is a select of i1 operands, we'll pattern match it.
N->getValueType(0) == MVT::i32) {
SDNode *Tmp =
CurDAG->getMachineNode(PPC::ADDIC, dl, MVT::i32, MVT::Glue,
- N->getOperand(0), getI32Imm(~0U));
+ N->getOperand(0), getI32Imm(~0U, dl));
return CurDAG->SelectNodeTo(N, PPC::SUBFE, MVT::i32,
SDValue(Tmp, 0), N->getOperand(0),
SDValue(Tmp, 1));
else if (N->getValueType(0) == MVT::i64)
SelectCCOp = PPC::SELECT_CC_I8;
else if (N->getValueType(0) == MVT::f32)
- SelectCCOp = PPC::SELECT_CC_F4;
+ if (PPCSubTarget->hasP8Vector())
+ SelectCCOp = PPC::SELECT_CC_VSSRC;
+ else
+ SelectCCOp = PPC::SELECT_CC_F4;
else if (N->getValueType(0) == MVT::f64)
if (PPCSubTarget->hasVSX())
SelectCCOp = PPC::SELECT_CC_VSFRC;
else
SelectCCOp = PPC::SELECT_CC_F8;
+ else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f64)
+ SelectCCOp = PPC::SELECT_CC_QFRC;
+ else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4f32)
+ SelectCCOp = PPC::SELECT_CC_QSRC;
+ else if (PPCSubTarget->hasQPX() && N->getValueType(0) == MVT::v4i1)
+ SelectCCOp = PPC::SELECT_CC_QBRC;
else if (N->getValueType(0) == MVT::v2f64 ||
N->getValueType(0) == MVT::v2i64)
SelectCCOp = PPC::SELECT_CC_VSRC;
SelectCCOp = PPC::SELECT_CC_VRRC;
SDValue Ops[] = { CCReg, N->getOperand(2), N->getOperand(3),
- getI32Imm(BROpc) };
+ getI32Imm(BROpc, dl) };
return CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), Ops);
}
case ISD::VSELECT:
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];
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() &&
+ if (LD->isUnindexed() && LD->hasOneUse() && Op1.hasOneUse() &&
+ (LD->getMemoryVT() == MVT::f64 ||
+ LD->getMemoryVT() == MVT::i64) &&
SelectAddrIdxOnly(LD->getBasePtr(), Base, Offset)) {
SDValue Chain = LD->getChain();
SDValue Ops[] = { Base, Offset, Chain };
}
}
+ // 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), dl,
+ MVT::i32);
SDValue Ops[] = { Op1, Op2, DMV };
return CurDAG->SelectNodeTo(N, PPC::XXPERMDI, N->getValueType(0), Ops);
}
// Op #4 is the Flag.
// Prevent PPC::PRED_* from being selected into LI.
SDValue Pred =
- getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue());
+ getI32Imm(cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(), dl);
SDValue Ops[] = { Pred, N->getOperand(2), N->getOperand(3),
N->getOperand(0), N->getOperand(4) };
return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
}
SDValue CondCode = SelectCC(N->getOperand(2), N->getOperand(3), CC, dl);
- SDValue Ops[] = { getI32Imm(PCC), CondCode,
+ SDValue Ops[] = { getI32Imm(PCC, dl), CondCode,
N->getOperand(4), N->getOperand(0) };
return CurDAG->SelectNodeTo(N, PPC::BCC, MVT::Other, Ops);
}
"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));
+ return transferMemOperands(N, CurDAG->getMachineNode(PPC::LWZtoc, dl,
+ MVT::i32, GA, N->getOperand(1)));
}
// For medium and large code model, we generate two instructions as
break;
// 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:
+ // If it must be toc-referenced according to PPCSubTarget, 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);
+ TOCbase, GA);
if (isa<JumpTableSDNode>(GA) || isa<BlockAddressSDNode>(GA) ||
CModel == CodeModel::Large)
- return CurDAG->getMachineNode(PPC::LDtocL, dl, MVT::i64, GA,
- SDValue(Tmp, 0));
+ return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl,
+ MVT::i64, GA, SDValue(Tmp, 0)));
if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(GA)) {
- const GlobalValue *GValue = G->getGlobal();
- 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));
+ const GlobalValue *GV = G->getGlobal();
+ unsigned char GVFlags = PPCSubTarget->classifyGlobalReference(GV);
+ if (GVFlags & PPCII::MO_NLP_FLAG) {
+ return transferMemOperands(N, CurDAG->getMachineNode(PPC::LDtocL, dl,
+ MVT::i64, GA, SDValue(Tmp, 0)));
+ }
}
return CurDAG->getMachineNode(PPC::ADDItocL, dl, MVT::i64,
// 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);
+ return CurDAG->SelectNodeTo(
+ N, PPC::PPC32PICGOT, PPCLowering->getPointerTy(CurDAG->getDataLayout()),
+ MVT::i32);
}
case PPCISD::VADD_SPLAT: {
// This expands into one of three sequences, depending on whether
// 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);
+ SDValue EltVal = getI32Imm(Elt >> 1, dl);
SDNode *Tmp = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
SDValue TmpVal = SDValue(Tmp, 0);
return CurDAG->getMachineNode(Opc2, dl, VT, TmpVal, TmpVal);
// Into: tmp1 = VSPLTIS[BHW] elt-16
// tmp2 = VSPLTIS[BHW] -16
// VSUBU[BHW]M tmp1, tmp2
- SDValue EltVal = getI32Imm(Elt - 16);
+ SDValue EltVal = getI32Imm(Elt - 16, dl);
SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
- EltVal = getI32Imm(-16);
+ EltVal = getI32Imm(-16, dl);
SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
return CurDAG->getMachineNode(Opc3, dl, VT, SDValue(Tmp1, 0),
SDValue(Tmp2, 0));
// Into: tmp1 = VSPLTIS[BHW] elt+16
// tmp2 = VSPLTIS[BHW] -16
// VADDU[BHW]M tmp1, tmp2
- SDValue EltVal = getI32Imm(Elt + 16);
+ SDValue EltVal = getI32Imm(Elt + 16, dl);
SDNode *Tmp1 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
- EltVal = getI32Imm(-16);
+ EltVal = getI32Imm(-16, dl);
SDNode *Tmp2 = CurDAG->getMachineNode(Opc1, dl, VT, EltVal);
return CurDAG->getMachineNode(Opc2, dl, VT, SDValue(Tmp1, 0),
SDValue(Tmp2, 0));
return SelectCode(N);
}
+// If the target supports the cmpb instruction, do the idiom recognition here.
+// We don't do this as a DAG combine because we don't want to do it as nodes
+// are being combined (because we might miss part of the eventual idiom). We
+// don't want to do it during instruction selection because we want to reuse
+// the logic for lowering the masking operations already part of the
+// instruction selector.
+SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) {
+ SDLoc dl(N);
+
+ assert(N->getOpcode() == ISD::OR &&
+ "Only OR nodes are supported for CMPB");
+
+ SDValue Res;
+ if (!PPCSubTarget->hasCMPB())
+ return Res;
+
+ if (N->getValueType(0) != MVT::i32 &&
+ N->getValueType(0) != MVT::i64)
+ return Res;
+
+ EVT VT = N->getValueType(0);
+
+ SDValue RHS, LHS;
+ bool BytesFound[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
+ uint64_t Mask = 0, Alt = 0;
+
+ auto IsByteSelectCC = [this](SDValue O, unsigned &b,
+ uint64_t &Mask, uint64_t &Alt,
+ SDValue &LHS, SDValue &RHS) {
+ if (O.getOpcode() != ISD::SELECT_CC)
+ return false;
+ ISD::CondCode CC = cast<CondCodeSDNode>(O.getOperand(4))->get();
+
+ if (!isa<ConstantSDNode>(O.getOperand(2)) ||
+ !isa<ConstantSDNode>(O.getOperand(3)))
+ return false;
+
+ uint64_t PM = O.getConstantOperandVal(2);
+ uint64_t PAlt = O.getConstantOperandVal(3);
+ for (b = 0; b < 8; ++b) {
+ uint64_t Mask = UINT64_C(0xFF) << (8*b);
+ if (PM && (PM & Mask) == PM && (PAlt & Mask) == PAlt)
+ break;
+ }
+
+ if (b == 8)
+ return false;
+ Mask |= PM;
+ Alt |= PAlt;
+
+ if (!isa<ConstantSDNode>(O.getOperand(1)) ||
+ O.getConstantOperandVal(1) != 0) {
+ SDValue Op0 = O.getOperand(0), Op1 = O.getOperand(1);
+ if (Op0.getOpcode() == ISD::TRUNCATE)
+ Op0 = Op0.getOperand(0);
+ if (Op1.getOpcode() == ISD::TRUNCATE)
+ Op1 = Op1.getOperand(0);
+
+ if (Op0.getOpcode() == ISD::SRL && Op1.getOpcode() == ISD::SRL &&
+ Op0.getOperand(1) == Op1.getOperand(1) && CC == ISD::SETEQ &&
+ isa<ConstantSDNode>(Op0.getOperand(1))) {
+
+ unsigned Bits = Op0.getValueType().getSizeInBits();
+ if (b != Bits/8-1)
+ return false;
+ if (Op0.getConstantOperandVal(1) != Bits-8)
+ return false;
+
+ LHS = Op0.getOperand(0);
+ RHS = Op1.getOperand(0);
+ return true;
+ }
+
+ // When we have small integers (i16 to be specific), the form present
+ // post-legalization uses SETULT in the SELECT_CC for the
+ // higher-order byte, depending on the fact that the
+ // even-higher-order bytes are known to all be zero, for example:
+ // select_cc (xor $lhs, $rhs), 256, 65280, 0, setult
+ // (so when the second byte is the same, because all higher-order
+ // bits from bytes 3 and 4 are known to be zero, the result of the
+ // xor can be at most 255)
+ if (Op0.getOpcode() == ISD::XOR && CC == ISD::SETULT &&
+ isa<ConstantSDNode>(O.getOperand(1))) {
+
+ uint64_t ULim = O.getConstantOperandVal(1);
+ if (ULim != (UINT64_C(1) << b*8))
+ return false;
+
+ // Now we need to make sure that the upper bytes are known to be
+ // zero.
+ unsigned Bits = Op0.getValueType().getSizeInBits();
+ if (!CurDAG->MaskedValueIsZero(Op0,
+ APInt::getHighBitsSet(Bits, Bits - (b+1)*8)))
+ return false;
+
+ LHS = Op0.getOperand(0);
+ RHS = Op0.getOperand(1);
+ return true;
+ }
+
+ return false;
+ }
+
+ if (CC != ISD::SETEQ)
+ return false;
+
+ SDValue Op = O.getOperand(0);
+ if (Op.getOpcode() == ISD::AND) {
+ if (!isa<ConstantSDNode>(Op.getOperand(1)))
+ return false;
+ if (Op.getConstantOperandVal(1) != (UINT64_C(0xFF) << (8*b)))
+ return false;
+
+ SDValue XOR = Op.getOperand(0);
+ if (XOR.getOpcode() == ISD::TRUNCATE)
+ XOR = XOR.getOperand(0);
+ if (XOR.getOpcode() != ISD::XOR)
+ return false;
+
+ LHS = XOR.getOperand(0);
+ RHS = XOR.getOperand(1);
+ return true;
+ } else if (Op.getOpcode() == ISD::SRL) {
+ if (!isa<ConstantSDNode>(Op.getOperand(1)))
+ return false;
+ unsigned Bits = Op.getValueType().getSizeInBits();
+ if (b != Bits/8-1)
+ return false;
+ if (Op.getConstantOperandVal(1) != Bits-8)
+ return false;
+
+ SDValue XOR = Op.getOperand(0);
+ if (XOR.getOpcode() == ISD::TRUNCATE)
+ XOR = XOR.getOperand(0);
+ if (XOR.getOpcode() != ISD::XOR)
+ return false;
+
+ LHS = XOR.getOperand(0);
+ RHS = XOR.getOperand(1);
+ return true;
+ }
+
+ return false;
+ };
+
+ SmallVector<SDValue, 8> Queue(1, SDValue(N, 0));
+ while (!Queue.empty()) {
+ SDValue V = Queue.pop_back_val();
+
+ for (const SDValue &O : V.getNode()->ops()) {
+ unsigned b;
+ uint64_t M = 0, A = 0;
+ SDValue OLHS, ORHS;
+ if (O.getOpcode() == ISD::OR) {
+ Queue.push_back(O);
+ } else if (IsByteSelectCC(O, b, M, A, OLHS, ORHS)) {
+ if (!LHS) {
+ LHS = OLHS;
+ RHS = ORHS;
+ BytesFound[b] = true;
+ Mask |= M;
+ Alt |= A;
+ } else if ((LHS == ORHS && RHS == OLHS) ||
+ (RHS == ORHS && LHS == OLHS)) {
+ BytesFound[b] = true;
+ Mask |= M;
+ Alt |= A;
+ } else {
+ return Res;
+ }
+ } else {
+ return Res;
+ }
+ }
+ }
+
+ unsigned LastB = 0, BCnt = 0;
+ for (unsigned i = 0; i < 8; ++i)
+ if (BytesFound[LastB]) {
+ ++BCnt;
+ LastB = i;
+ }
+
+ if (!LastB || BCnt < 2)
+ return Res;
+
+ // Because we'll be zero-extending the output anyway if don't have a specific
+ // value for each input byte (via the Mask), we can 'anyext' the inputs.
+ if (LHS.getValueType() != VT) {
+ LHS = CurDAG->getAnyExtOrTrunc(LHS, dl, VT);
+ RHS = CurDAG->getAnyExtOrTrunc(RHS, dl, VT);
+ }
+
+ Res = CurDAG->getNode(PPCISD::CMPB, dl, VT, LHS, RHS);
+
+ bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1);
+ if (NonTrivialMask && !Alt) {
+ // Res = Mask & CMPB
+ Res = CurDAG->getNode(ISD::AND, dl, VT, Res,
+ CurDAG->getConstant(Mask, dl, VT));
+ } else if (Alt) {
+ // Res = (CMPB & Mask) | (~CMPB & Alt)
+ // Which, as suggested here:
+ // https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge
+ // can be written as:
+ // Res = Alt ^ ((Alt ^ Mask) & CMPB)
+ // useful because the (Alt ^ Mask) can be pre-computed.
+ Res = CurDAG->getNode(ISD::AND, dl, VT, Res,
+ CurDAG->getConstant(Mask ^ Alt, dl, VT));
+ Res = CurDAG->getNode(ISD::XOR, dl, VT, Res,
+ CurDAG->getConstant(Alt, dl, VT));
+ }
+
+ return Res;
+}
+
+// When CR bit registers are enabled, an extension of an i1 variable to a i32
+// or i64 value is lowered in terms of a SELECT_I[48] operation, and thus
+// involves constant materialization of a 0 or a 1 or both. If the result of
+// the extension is then operated upon by some operator that can be constant
+// folded with a constant 0 or 1, and that constant can be materialized using
+// only one instruction (like a zero or one), then we should fold in those
+// operations with the select.
+void PPCDAGToDAGISel::foldBoolExts(SDValue &Res, SDNode *&N) {
+ if (!PPCSubTarget->useCRBits())
+ return;
+
+ if (N->getOpcode() != ISD::ZERO_EXTEND &&
+ N->getOpcode() != ISD::SIGN_EXTEND &&
+ N->getOpcode() != ISD::ANY_EXTEND)
+ return;
+
+ if (N->getOperand(0).getValueType() != MVT::i1)
+ return;
+
+ if (!N->hasOneUse())
+ return;
+
+ SDLoc dl(N);
+ EVT VT = N->getValueType(0);
+ SDValue Cond = N->getOperand(0);
+ SDValue ConstTrue =
+ CurDAG->getConstant(N->getOpcode() == ISD::SIGN_EXTEND ? -1 : 1, dl, VT);
+ SDValue ConstFalse = CurDAG->getConstant(0, dl, VT);
+
+ do {
+ SDNode *User = *N->use_begin();
+ if (User->getNumOperands() != 2)
+ break;
+
+ auto TryFold = [this, N, User, dl](SDValue Val) {
+ SDValue UserO0 = User->getOperand(0), UserO1 = User->getOperand(1);
+ SDValue O0 = UserO0.getNode() == N ? Val : UserO0;
+ SDValue O1 = UserO1.getNode() == N ? Val : UserO1;
+
+ return CurDAG->FoldConstantArithmetic(User->getOpcode(), dl,
+ User->getValueType(0),
+ O0.getNode(), O1.getNode());
+ };
+
+ SDValue TrueRes = TryFold(ConstTrue);
+ if (!TrueRes)
+ break;
+ SDValue FalseRes = TryFold(ConstFalse);
+ if (!FalseRes)
+ break;
+
+ // For us to materialize these using one instruction, we must be able to
+ // represent them as signed 16-bit integers.
+ uint64_t True = cast<ConstantSDNode>(TrueRes)->getZExtValue(),
+ False = cast<ConstantSDNode>(FalseRes)->getZExtValue();
+ if (!isInt<16>(True) || !isInt<16>(False))
+ break;
+
+ // We can replace User with a new SELECT node, and try again to see if we
+ // can fold the select with its user.
+ Res = CurDAG->getSelect(dl, User->getValueType(0), Cond, TrueRes, FalseRes);
+ N = User;
+ ConstTrue = TrueRes;
+ ConstFalse = FalseRes;
+ } while (N->hasOneUse());
+}
+
+void PPCDAGToDAGISel::PreprocessISelDAG() {
+ SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
+ ++Position;
+
+ bool MadeChange = false;
+ while (Position != CurDAG->allnodes_begin()) {
+ SDNode *N = &*--Position;
+ if (N->use_empty())
+ continue;
+
+ SDValue Res;
+ switch (N->getOpcode()) {
+ default: break;
+ case ISD::OR:
+ Res = combineToCMPB(N);
+ break;
+ }
+
+ if (!Res)
+ foldBoolExts(Res, N);
+
+ if (Res) {
+ DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld: ");
+ DEBUG(N->dump(CurDAG));
+ DEBUG(dbgs() << "\nNew: ");
+ DEBUG(Res.getNode()->dump(CurDAG));
+ DEBUG(dbgs() << "\n");
+
+ CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), Res);
+ MadeChange = true;
+ }
+ }
+
+ if (MadeChange)
+ CurDAG->RemoveDeadNodes();
+}
+
/// PostprocessISelDAG - Perform some late peephole optimizations
/// on the DAG representation.
void PPCDAGToDAGISel::PostprocessISelDAG() {
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);
+ for (SDNode &Node : CurDAG->allnodes()) {
+ MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(&Node);
if (!MachineNode || MachineNode->use_empty())
continue;
SDNode *ResNode = MachineNode;
case PPC::SELECT_I8:
case PPC::SELECT_F4:
case PPC::SELECT_F8:
+ case PPC::SELECT_QFRC:
+ case PPC::SELECT_QSRC:
+ case PPC::SELECT_QBRC:
case PPC::SELECT_VRRC:
case PPC::SELECT_VSFRC:
+ case PPC::SELECT_VSSRC:
case PPC::SELECT_VSRC: {
SDValue Op = MachineNode->getOperand(0);
if (Op.isMachineOpcode()) {
case PPC::SELECT_I8:
case PPC::SELECT_F4:
case PPC::SELECT_F8:
+ case PPC::SELECT_QFRC:
+ case PPC::SELECT_QSRC:
+ case PPC::SELECT_QBRC:
case PPC::SELECT_VRRC:
case PPC::SELECT_VSFRC:
+ case PPC::SELECT_VSSRC:
case PPC::SELECT_VSRC:
if (Op1Set)
ResNode = MachineNode->getOperand(1).getNode();
return true;
}
+ // LHBRX and LWBRX always clear the higher-order bits.
+ if (Op32.getMachineOpcode() == PPC::LHBRX ||
+ Op32.getMachineOpcode() == PPC::LWBRX) {
+ ToPromote.insert(Op32.getNode());
+ return true;
+ }
+
+ // CNTLZW always produces a 64-bit value in [0,32], and so is zero extended.
+ if (Op32.getMachineOpcode() == PPC::CNTLZW) {
+ 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
bool MadeChange = false;
while (Position != CurDAG->allnodes_begin()) {
- SDNode *N = --Position;
+ SDNode *N = &*--Position;
// Skip dead nodes and any non-machine opcodes.
if (N->use_empty() || !N->isMachineOpcode())
continue;
case PPC::SRW: NewOpcode = PPC::SRW8; break;
case PPC::LI: NewOpcode = PPC::LI8; break;
case PPC::LIS: NewOpcode = PPC::LIS8; break;
+ case PPC::LHBRX: NewOpcode = PPC::LHBRX8; break;
+ case PPC::LWBRX: NewOpcode = PPC::LWBRX8; break;
+ case PPC::CNTLZW: NewOpcode = PPC::CNTLZW8; break;
case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break;
case PPC::OR: NewOpcode = PPC::OR8; break;
case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break;
++Position;
while (Position != CurDAG->allnodes_begin()) {
- SDNode *N = --Position;
+ SDNode *N = &*--Position;
// Skip dead nodes and any non-machine opcodes.
if (N->use_empty() || !N->isMachineOpcode())
continue;
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)
+ // If this is a load or store with a zero offset, or within the alignment,
+ // we may be able to fold an add-immediate into the memory operation.
+ // The check against alignment is below, as it can't occur until we check
+ // the arguments to N
+ if (!isa<ConstantSDNode>(N->getOperand(FirstOp)))
continue;
SDValue Base = N->getOperand(FirstOp + 1);
if (!Base.isMachineOpcode())
continue;
+ // On targets with fusion, we don't want this to fire and remove a fusion
+ // opportunity, unless a) it results in another fusion opportunity or
+ // b) optimizing for size.
+ if (PPCSubTarget->hasFusion() &&
+ (!MF->getFunction()->optForSize() && !Base.hasOneUse()))
+ continue;
+
unsigned Flags = 0;
bool ReplaceFlags = true;
break;
}
+ SDValue ImmOpnd = Base.getOperand(1);
+ int MaxDisplacement = 0;
+ if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
+ const GlobalValue *GV = GA->getGlobal();
+ MaxDisplacement = GV->getAlignment() - 1;
+ }
+
+ int Offset = N->getConstantOperandVal(FirstOp);
+ if (Offset < 0 || Offset > MaxDisplacement)
+ continue;
+
// 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
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) {
// is insufficient for the instruction encoding.
if (GV->getAlignment() < 4 &&
(StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
- StorageOpcode == PPC::LWA)) {
+ StorageOpcode == PPC::LWA || (Offset % 4) != 0)) {
DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
continue;
}
- ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags);
+ ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, Offset, Flags);
} else if (ConstantPoolSDNode *CP =
dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
const Constant *C = CP->getConstVal();
ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
CP->getAlignment(),
- 0, Flags);
+ Offset, Flags);
}
}
projects / oota-llvm.git / blobdiff