+MachineBasicBlock*
+SparcTargetLowering::expandAtomicRMW(MachineInstr *MI,
+ MachineBasicBlock *MBB,
+ unsigned Opcode,
+ unsigned CondCode) const {
+ const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
+ MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
+ DebugLoc DL = MI->getDebugLoc();
+
+ // MI is an atomic read-modify-write instruction of the form:
+ //
+ // rd = atomicrmw<op> addr, rs2
+ //
+ // All three operands are registers.
+ unsigned DestReg = MI->getOperand(0).getReg();
+ unsigned AddrReg = MI->getOperand(1).getReg();
+ unsigned Rs2Reg = MI->getOperand(2).getReg();
+
+ // SelectionDAG has already inserted memory barriers before and after MI, so
+ // we simply have to implement the operatiuon in terms of compare-and-swap.
+ //
+ // %val0 = load %addr
+ // loop:
+ // %val = phi %val0, %dest
+ // %upd = op %val, %rs2
+ // %dest = cas %addr, %val, %upd
+ // cmp %val, %dest
+ // bne loop
+ // done:
+ //
+ bool is64Bit = SP::I64RegsRegClass.hasSubClassEq(MRI.getRegClass(DestReg));
+ const TargetRegisterClass *ValueRC =
+ is64Bit ? &SP::I64RegsRegClass : &SP::IntRegsRegClass;
+ unsigned Val0Reg = MRI.createVirtualRegister(ValueRC);
+
+ BuildMI(*MBB, MI, DL, TII.get(is64Bit ? SP::LDXri : SP::LDri), Val0Reg)
+ .addReg(AddrReg).addImm(0);
+
+ // Split the basic block MBB before MI and insert the loop block in the hole.
+ MachineFunction::iterator MFI = MBB->getIterator();
+ const BasicBlock *LLVM_BB = MBB->getBasicBlock();
+ MachineFunction *MF = MBB->getParent();
+ MachineBasicBlock *LoopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *DoneMBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ ++MFI;
+ MF->insert(MFI, LoopMBB);
+ MF->insert(MFI, DoneMBB);
+
+ // Move MI and following instructions to DoneMBB.
+ DoneMBB->splice(DoneMBB->begin(), MBB, MI, MBB->end());
+ DoneMBB->transferSuccessorsAndUpdatePHIs(MBB);
+
+ // Connect the CFG again.
+ MBB->addSuccessor(LoopMBB);
+ LoopMBB->addSuccessor(LoopMBB);
+ LoopMBB->addSuccessor(DoneMBB);
+
+ // Build the loop block.
+ unsigned ValReg = MRI.createVirtualRegister(ValueRC);
+ // Opcode == 0 means try to write Rs2Reg directly (ATOMIC_SWAP).
+ unsigned UpdReg = (Opcode ? MRI.createVirtualRegister(ValueRC) : Rs2Reg);
+
+ BuildMI(LoopMBB, DL, TII.get(SP::PHI), ValReg)
+ .addReg(Val0Reg).addMBB(MBB)
+ .addReg(DestReg).addMBB(LoopMBB);
+
+ if (CondCode) {
+ // This is one of the min/max operations. We need a CMPrr followed by a
+ // MOVXCC/MOVICC.
+ BuildMI(LoopMBB, DL, TII.get(SP::CMPrr)).addReg(ValReg).addReg(Rs2Reg);
+ BuildMI(LoopMBB, DL, TII.get(Opcode), UpdReg)
+ .addReg(ValReg).addReg(Rs2Reg).addImm(CondCode);
+ } else if (Opcode) {
+ BuildMI(LoopMBB, DL, TII.get(Opcode), UpdReg)
+ .addReg(ValReg).addReg(Rs2Reg);
+ }
+
+ if (MI->getOpcode() == SP::ATOMIC_LOAD_NAND_32 ||
+ MI->getOpcode() == SP::ATOMIC_LOAD_NAND_64) {
+ unsigned TmpReg = UpdReg;
+ UpdReg = MRI.createVirtualRegister(ValueRC);
+ BuildMI(LoopMBB, DL, TII.get(SP::XORri), UpdReg).addReg(TmpReg).addImm(-1);
+ }
+
+ BuildMI(LoopMBB, DL, TII.get(is64Bit ? SP::CASXrr : SP::CASrr), DestReg)
+ .addReg(AddrReg).addReg(ValReg).addReg(UpdReg)
+ .setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
+ BuildMI(LoopMBB, DL, TII.get(SP::CMPrr)).addReg(ValReg).addReg(DestReg);
+ BuildMI(LoopMBB, DL, TII.get(is64Bit ? SP::BPXCC : SP::BCOND))
+ .addMBB(LoopMBB).addImm(SPCC::ICC_NE);
+
+ MI->eraseFromParent();
+ return DoneMBB;
+}
+