#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/raw_ostream.h"
#include <sstream>
using namespace llvm;
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
else
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
- setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
+ setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
+ case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
+ case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
+
case ARMISD::VCEQ: return "ARMISD::VCEQ";
case ARMISD::VCGE: return "ARMISD::VCGE";
case ARMISD::VCGEU: return "ARMISD::VCGEU";
LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
false, false, false, false,
0, CallingConv::C, false, /*isReturnValueUsed=*/true,
- DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
+ DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl,
+ DAG.GetOrdering(Chain.getNode()));
return CallResult.first;
}
}
}
+static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
+ const ARMSubtarget *Subtarget) {
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue Op5 = Op.getOperand(5);
+ SDValue Res;
+ unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
+ if (isDeviceBarrier) {
+ if (Subtarget->hasV7Ops())
+ Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
+ else
+ Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
+ DAG.getConstant(0, MVT::i32));
+ } else {
+ if (Subtarget->hasV7Ops())
+ Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
+ else
+ Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
+ DAG.getConstant(0, MVT::i32));
+ }
+ return Res;
+}
+
static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
unsigned VarArgsFrameIndex) {
// vastart just stores the address of the VarArgsFrameIndex slot into the
return true;
}
+/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
+static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
+ WhichResult = (M[0] == 0 ? 0 : 1);
+ for (unsigned i = 0; i < NumElts; i += 2) {
+ if ((unsigned) M[i] != i + WhichResult ||
+ (unsigned) M[i+1] != i + WhichResult)
+ return false;
+ }
+ return true;
+}
+
static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
unsigned &WhichResult) {
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
return true;
}
+/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
+static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned Half = VT.getVectorNumElements() / 2;
+ WhichResult = (M[0] == 0 ? 0 : 1);
+ for (unsigned j = 0; j != 2; ++j) {
+ unsigned Idx = WhichResult;
+ for (unsigned i = 0; i != Half; ++i) {
+ if ((unsigned) M[i + j * Half] != Idx)
+ return false;
+ Idx += 2;
+ }
+ }
+
+ // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+ if (VT.is64BitVector() && EltSz == 32)
+ return false;
+
+ return true;
+}
+
static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
unsigned &WhichResult) {
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
return true;
}
+/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
+static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
+ WhichResult = (M[0] == 0 ? 0 : 1);
+ unsigned Idx = WhichResult * NumElts / 2;
+ for (unsigned i = 0; i != NumElts; i += 2) {
+ if ((unsigned) M[i] != Idx ||
+ (unsigned) M[i+1] != Idx)
+ return false;
+ Idx += 1;
+ }
+
+ // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+ if (VT.is64BitVector() && EltSz == 32)
+ return false;
+
+ return true;
+}
+
+
static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
// Canonicalize all-zeros and all-ones vectors.
ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
isVEXTMask(M, VT, ReverseVEXT, Imm) ||
isVTRNMask(M, VT, WhichResult) ||
isVUZPMask(M, VT, WhichResult) ||
- isVZIPMask(M, VT, WhichResult));
+ isVZIPMask(M, VT, WhichResult) ||
+ isVTRN_v_undef_Mask(M, VT, WhichResult) ||
+ isVUZP_v_undef_Mask(M, VT, WhichResult) ||
+ isVZIP_v_undef_Mask(M, VT, WhichResult));
}
/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
V1, V2).getValue(WhichResult);
+ if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
+ V1, V1).getValue(WhichResult);
+ if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
+ V1, V1).getValue(WhichResult);
+ if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
+ V1, V1).getValue(WhichResult);
+
// If the shuffle is not directly supported and it has 4 elements, use
// the PerfectShuffle-generated table to synthesize it from other shuffles.
if (VT.getVectorNumElements() == 4 &&
case ISD::BR_JT: return LowerBR_JT(Op, DAG);
case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
+ case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
case ISD::SINT_TO_FP:
case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
case ISD::FP_TO_SINT:
// ARM Scheduler Hooks
//===----------------------------------------------------------------------===//
+MachineBasicBlock *
+ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
+ MachineBasicBlock *BB,
+ unsigned Size) const {
+ unsigned dest = MI->getOperand(0).getReg();
+ unsigned ptr = MI->getOperand(1).getReg();
+ unsigned oldval = MI->getOperand(2).getReg();
+ unsigned newval = MI->getOperand(3).getReg();
+ unsigned scratch = BB->getParent()->getRegInfo()
+ .createVirtualRegister(ARM::GPRRegisterClass);
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ DebugLoc dl = MI->getDebugLoc();
+ bool isThumb2 = Subtarget->isThumb2();
+
+ unsigned ldrOpc, strOpc;
+ switch (Size) {
+ default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
+ case 1:
+ ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
+ strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
+ break;
+ case 2:
+ ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
+ strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
+ break;
+ case 4:
+ ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
+ strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
+ break;
+ }
+
+ MachineFunction *MF = BB->getParent();
+ const BasicBlock *LLVM_BB = BB->getBasicBlock();
+ MachineFunction::iterator It = BB;
+ ++It; // insert the new blocks after the current block
+
+ MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ MF->insert(It, loop1MBB);
+ MF->insert(It, loop2MBB);
+ MF->insert(It, exitMBB);
+ exitMBB->transferSuccessors(BB);
+
+ // thisMBB:
+ // ...
+ // fallthrough --> loop1MBB
+ BB->addSuccessor(loop1MBB);
+
+ // loop1MBB:
+ // ldrex dest, [ptr]
+ // cmp dest, oldval
+ // bne exitMBB
+ BB = loop1MBB;
+ AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
+ AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
+ .addReg(dest).addReg(oldval));
+ BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
+ .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
+ BB->addSuccessor(loop2MBB);
+ BB->addSuccessor(exitMBB);
+
+ // loop2MBB:
+ // strex scratch, newval, [ptr]
+ // cmp scratch, #0
+ // bne loop1MBB
+ BB = loop2MBB;
+ AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
+ .addReg(ptr));
+ AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
+ .addReg(scratch).addImm(0));
+ BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
+ .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
+ BB->addSuccessor(loop1MBB);
+ BB->addSuccessor(exitMBB);
+
+ // exitMBB:
+ // ...
+ BB = exitMBB;
+
+ MF->DeleteMachineInstr(MI); // The instruction is gone now.
+
+ return BB;
+}
+
+MachineBasicBlock *
+ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
+ unsigned Size, unsigned BinOpcode) const {
+ // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+
+ const BasicBlock *LLVM_BB = BB->getBasicBlock();
+ MachineFunction *F = BB->getParent();
+ MachineFunction::iterator It = BB;
+ ++It;
+
+ unsigned dest = MI->getOperand(0).getReg();
+ unsigned ptr = MI->getOperand(1).getReg();
+ unsigned incr = MI->getOperand(2).getReg();
+ DebugLoc dl = MI->getDebugLoc();
+
+ bool isThumb2 = Subtarget->isThumb2();
+ unsigned ldrOpc, strOpc;
+ switch (Size) {
+ default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
+ case 1:
+ ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
+ strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
+ break;
+ case 2:
+ ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
+ strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
+ break;
+ case 4:
+ ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
+ strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
+ break;
+ }
+
+ MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
+ F->insert(It, loopMBB);
+ F->insert(It, exitMBB);
+ exitMBB->transferSuccessors(BB);
+
+ MachineRegisterInfo &RegInfo = F->getRegInfo();
+ unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
+ unsigned scratch2 = (!BinOpcode) ? incr :
+ RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
+
+ // thisMBB:
+ // ...
+ // fallthrough --> loopMBB
+ BB->addSuccessor(loopMBB);
+
+ // loopMBB:
+ // ldrex dest, ptr
+ // <binop> scratch2, dest, incr
+ // strex scratch, scratch2, ptr
+ // cmp scratch, #0
+ // bne- loopMBB
+ // fallthrough --> exitMBB
+ BB = loopMBB;
+ AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
+ if (BinOpcode) {
+ // operand order needs to go the other way for NAND
+ if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
+ AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
+ addReg(incr).addReg(dest)).addReg(0);
+ else
+ AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
+ addReg(dest).addReg(incr)).addReg(0);
+ }
+
+ AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
+ .addReg(ptr));
+ AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
+ .addReg(scratch).addImm(0));
+ BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
+ .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
+
+ BB->addSuccessor(loopMBB);
+ BB->addSuccessor(exitMBB);
+
+ // exitMBB:
+ // ...
+ BB = exitMBB;
+
+ F->DeleteMachineInstr(MI); // The instruction is gone now.
+
+ return BB;
+}
+
MachineBasicBlock *
ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
MachineBasicBlock *BB,
DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
DebugLoc dl = MI->getDebugLoc();
+ bool isThumb2 = Subtarget->isThumb2();
switch (MI->getOpcode()) {
default:
+ MI->dump();
llvm_unreachable("Unexpected instr type to insert");
+
+ case ARM::ATOMIC_LOAD_ADD_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
+ case ARM::ATOMIC_LOAD_ADD_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
+ case ARM::ATOMIC_LOAD_ADD_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
+
+ case ARM::ATOMIC_LOAD_AND_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
+ case ARM::ATOMIC_LOAD_AND_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
+ case ARM::ATOMIC_LOAD_AND_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
+
+ case ARM::ATOMIC_LOAD_OR_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
+ case ARM::ATOMIC_LOAD_OR_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
+ case ARM::ATOMIC_LOAD_OR_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
+
+ case ARM::ATOMIC_LOAD_XOR_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
+ case ARM::ATOMIC_LOAD_XOR_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
+ case ARM::ATOMIC_LOAD_XOR_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
+
+ case ARM::ATOMIC_LOAD_NAND_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
+ case ARM::ATOMIC_LOAD_NAND_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
+ case ARM::ATOMIC_LOAD_NAND_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
+
+ case ARM::ATOMIC_LOAD_SUB_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
+ case ARM::ATOMIC_LOAD_SUB_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
+ case ARM::ATOMIC_LOAD_SUB_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
+
+ case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
+ case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
+ case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
+
+ case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
+ case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
+ case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
+
case ARM::tMOVCCr_pseudo: {
// To "insert" a SELECT_CC instruction, we actually have to insert the
// diamond control-flow pattern. The incoming instruction knows the
ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
EVT VT) const {
if (Constraint.size() == 1) {
- // GCC RS6000 Constraint Letters
+ // GCC ARM Constraint Letters
switch (Constraint[0]) {
case 'l':
- if (Subtarget->isThumb1Only())
+ if (Subtarget->isThumb())
return std::make_pair(0U, ARM::tGPRRegisterClass);
else
return std::make_pair(0U, ARM::GPRRegisterClass);
case 'w':
if (VT == MVT::f32)
return std::make_pair(0U, ARM::SPRRegisterClass);
- if (VT == MVT::f64)
+ if (VT.getSizeInBits() == 64)
return std::make_pair(0U, ARM::DPRRegisterClass);
+ if (VT.getSizeInBits() == 128)
+ return std::make_pair(0U, ARM::QPRRegisterClass);
break;
}
}
ARM::S20,ARM::S21,ARM::S22,ARM::S23,
ARM::S24,ARM::S25,ARM::S26,ARM::S27,
ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
- if (VT == MVT::f64)
+ if (VT.getSizeInBits() == 64)
return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
ARM::D4, ARM::D5, ARM::D6, ARM::D7,
ARM::D8, ARM::D9, ARM::D10,ARM::D11,
ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
+ if (VT.getSizeInBits() == 128)
+ return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
+ ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
break;
}
projects / oota-llvm.git / blobdiff