+ unsigned DstReg, SrcReg, DReg;
+ unsigned Lane;
+ MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
+ const TargetRegisterInfo *TRI = &getRegisterInfo();
+ switch (MI->getOpcode()) {
+ default:
+ llvm_unreachable("cannot handle opcode!");
+ break;
+ case ARM::VMOVD:
+ if (Domain != ExeNEON)
+ break;
+
+ // Zap the predicate operands.
+ assert(!isPredicated(MI) && "Cannot predicate a VORRd");
+
+ // Source instruction is %DDst = VMOVD %DSrc, 14, %noreg (; implicits)
+ DstReg = MI->getOperand(0).getReg();
+ SrcReg = MI->getOperand(1).getReg();
+
+ for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
+ MI->RemoveOperand(i-1);
+
+ // Change to a %DDst = VORRd %DSrc, %DSrc, 14, %noreg (; implicits)
+ MI->setDesc(get(ARM::VORRd));
+ AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
+ .addReg(SrcReg)
+ .addReg(SrcReg));
+ break;
+ case ARM::VMOVRS:
+ if (Domain != ExeNEON)
+ break;
+ assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
+
+ // Source instruction is %RDst = VMOVRS %SSrc, 14, %noreg (; implicits)
+ DstReg = MI->getOperand(0).getReg();
+ SrcReg = MI->getOperand(1).getReg();
+
+ for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
+ MI->RemoveOperand(i-1);
+
+ DReg = getCorrespondingDRegAndLane(TRI, SrcReg, Lane);
+
+ // Convert to %RDst = VGETLNi32 %DSrc, Lane, 14, %noreg (; imps)
+ // Note that DSrc has been widened and the other lane may be undef, which
+ // contaminates the entire register.
+ MI->setDesc(get(ARM::VGETLNi32));
+ AddDefaultPred(MIB.addReg(DstReg, RegState::Define)
+ .addReg(DReg, RegState::Undef)
+ .addImm(Lane));
+
+ // The old source should be an implicit use, otherwise we might think it
+ // was dead before here.
+ MIB.addReg(SrcReg, RegState::Implicit);
+ break;
+ case ARM::VMOVSR: {
+ if (Domain != ExeNEON)
+ break;
+ assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
+
+ // Source instruction is %SDst = VMOVSR %RSrc, 14, %noreg (; implicits)
+ DstReg = MI->getOperand(0).getReg();
+ SrcReg = MI->getOperand(1).getReg();
+
+ DReg = getCorrespondingDRegAndLane(TRI, DstReg, Lane);
+
+ unsigned ImplicitSReg;
+ if (!getImplicitSPRUseForDPRUse(TRI, MI, DReg, Lane, ImplicitSReg))
+ break;
+
+ for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
+ MI->RemoveOperand(i-1);
+
+ // Convert to %DDst = VSETLNi32 %DDst, %RSrc, Lane, 14, %noreg (; imps)
+ // Again DDst may be undefined at the beginning of this instruction.
+ MI->setDesc(get(ARM::VSETLNi32));
+ MIB.addReg(DReg, RegState::Define)
+ .addReg(DReg, getUndefRegState(!MI->readsRegister(DReg, TRI)))
+ .addReg(SrcReg)
+ .addImm(Lane);
+ AddDefaultPred(MIB);
+
+ // The narrower destination must be marked as set to keep previous chains
+ // in place.
+ MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
+ if (ImplicitSReg != 0)
+ MIB.addReg(ImplicitSReg, RegState::Implicit);
+ break;
+ }
+ case ARM::VMOVS: {
+ if (Domain != ExeNEON)
+ break;
+
+ // Source instruction is %SDst = VMOVS %SSrc, 14, %noreg (; implicits)
+ DstReg = MI->getOperand(0).getReg();
+ SrcReg = MI->getOperand(1).getReg();
+
+ unsigned DstLane = 0, SrcLane = 0, DDst, DSrc;
+ DDst = getCorrespondingDRegAndLane(TRI, DstReg, DstLane);
+ DSrc = getCorrespondingDRegAndLane(TRI, SrcReg, SrcLane);
+
+ unsigned ImplicitSReg;
+ if (!getImplicitSPRUseForDPRUse(TRI, MI, DSrc, SrcLane, ImplicitSReg))
+ break;
+
+ for (unsigned i = MI->getDesc().getNumOperands(); i; --i)
+ MI->RemoveOperand(i-1);
+
+ if (DSrc == DDst) {
+ // Destination can be:
+ // %DDst = VDUPLN32d %DDst, Lane, 14, %noreg (; implicits)
+ MI->setDesc(get(ARM::VDUPLN32d));
+ MIB.addReg(DDst, RegState::Define)
+ .addReg(DDst, getUndefRegState(!MI->readsRegister(DDst, TRI)))
+ .addImm(SrcLane);
+ AddDefaultPred(MIB);
+
+ // Neither the source or the destination are naturally represented any
+ // more, so add them in manually.
+ MIB.addReg(DstReg, RegState::Implicit | RegState::Define);
+ MIB.addReg(SrcReg, RegState::Implicit);
+ if (ImplicitSReg != 0)
+ MIB.addReg(ImplicitSReg, RegState::Implicit);
+ break;
+ }
+
+ // In general there's no single instruction that can perform an S <-> S
+ // move in NEON space, but a pair of VEXT instructions *can* do the
+ // job. It turns out that the VEXTs needed will only use DSrc once, with
+ // the position based purely on the combination of lane-0 and lane-1
+ // involved. For example
+ // vmov s0, s2 -> vext.32 d0, d0, d1, #1 vext.32 d0, d0, d0, #1
+ // vmov s1, s3 -> vext.32 d0, d1, d0, #1 vext.32 d0, d0, d0, #1
+ // vmov s0, s3 -> vext.32 d0, d0, d0, #1 vext.32 d0, d1, d0, #1
+ // vmov s1, s2 -> vext.32 d0, d0, d0, #1 vext.32 d0, d0, d1, #1
+ //
+ // Pattern of the MachineInstrs is:
+ // %DDst = VEXTd32 %DSrc1, %DSrc2, Lane, 14, %noreg (;implicits)
+ MachineInstrBuilder NewMIB;
+ NewMIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
+ get(ARM::VEXTd32), DDst);
+
+ // On the first instruction, both DSrc and DDst may be <undef> if present.
+ // Specifically when the original instruction didn't have them as an
+ // <imp-use>.
+ unsigned CurReg = SrcLane == 1 && DstLane == 1 ? DSrc : DDst;
+ bool CurUndef = !MI->readsRegister(CurReg, TRI);
+ NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
+
+ CurReg = SrcLane == 0 && DstLane == 0 ? DSrc : DDst;
+ CurUndef = !MI->readsRegister(CurReg, TRI);
+ NewMIB.addReg(CurReg, getUndefRegState(CurUndef));
+
+ NewMIB.addImm(1);
+ AddDefaultPred(NewMIB);
+
+ if (SrcLane == DstLane)
+ NewMIB.addReg(SrcReg, RegState::Implicit);
+
+ MI->setDesc(get(ARM::VEXTd32));
+ MIB.addReg(DDst, RegState::Define);
+
+ // On the second instruction, DDst has definitely been defined above, so
+ // it is not <undef>. DSrc, if present, can be <undef> as above.
+ CurReg = SrcLane == 1 && DstLane == 0 ? DSrc : DDst;
+ CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
+ MIB.addReg(CurReg, getUndefRegState(CurUndef));
+
+ CurReg = SrcLane == 0 && DstLane == 1 ? DSrc : DDst;
+ CurUndef = CurReg == DSrc && !MI->readsRegister(CurReg, TRI);
+ MIB.addReg(CurReg, getUndefRegState(CurUndef));
+
+ MIB.addImm(1);
+ AddDefaultPred(MIB);
+
+ if (SrcLane != DstLane)
+ MIB.addReg(SrcReg, RegState::Implicit);
+
+ // As before, the original destination is no longer represented, add it
+ // implicitly.
+ MIB.addReg(DstReg, RegState::Define | RegState::Implicit);
+ if (ImplicitSReg != 0)
+ MIB.addReg(ImplicitSReg, RegState::Implicit);
+ break;
+ }
+ }
+
+}
+
+//===----------------------------------------------------------------------===//
+// Partial register updates
+//===----------------------------------------------------------------------===//
+//
+// Swift renames NEON registers with 64-bit granularity. That means any
+// instruction writing an S-reg implicitly reads the containing D-reg. The
+// problem is mostly avoided by translating f32 operations to v2f32 operations
+// on D-registers, but f32 loads are still a problem.
+//
+// These instructions can load an f32 into a NEON register:
+//
+// VLDRS - Only writes S, partial D update.
+// VLD1LNd32 - Writes all D-regs, explicit partial D update, 2 uops.
+// VLD1DUPd32 - Writes all D-regs, no partial reg update, 2 uops.
+//
+// FCONSTD can be used as a dependency-breaking instruction.
+
+
+unsigned ARMBaseInstrInfo::
+getPartialRegUpdateClearance(const MachineInstr *MI,
+ unsigned OpNum,
+ const TargetRegisterInfo *TRI) const {
+ // Only Swift has partial register update problems.
+ if (!SwiftPartialUpdateClearance || !Subtarget.isSwift())
+ return 0;
+
+ assert(TRI && "Need TRI instance");
+
+ const MachineOperand &MO = MI->getOperand(OpNum);
+ if (MO.readsReg())
+ return 0;
+ unsigned Reg = MO.getReg();
+ int UseOp = -1;
+
+ switch(MI->getOpcode()) {
+ // Normal instructions writing only an S-register.
+ case ARM::VLDRS:
+ case ARM::FCONSTS:
+ case ARM::VMOVSR:
+ case ARM::VMOVv8i8:
+ case ARM::VMOVv4i16:
+ case ARM::VMOVv2i32:
+ case ARM::VMOVv2f32:
+ case ARM::VMOVv1i64:
+ UseOp = MI->findRegisterUseOperandIdx(Reg, false, TRI);
+ break;
+
+ // Explicitly reads the dependency.
+ case ARM::VLD1LNd32:
+ UseOp = 1;
+ break;
+ default:
+ return 0;
+ }
+
+ // If this instruction actually reads a value from Reg, there is no unwanted
+ // dependency.
+ if (UseOp != -1 && MI->getOperand(UseOp).readsReg())
+ return 0;
+
+ // We must be able to clobber the whole D-reg.
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
+ // Virtual register must be a foo:ssub_0<def,undef> operand.
+ if (!MO.getSubReg() || MI->readsVirtualRegister(Reg))
+ return 0;
+ } else if (ARM::SPRRegClass.contains(Reg)) {
+ // Physical register: MI must define the full D-reg.
+ unsigned DReg = TRI->getMatchingSuperReg(Reg, ARM::ssub_0,
+ &ARM::DPRRegClass);
+ if (!DReg || !MI->definesRegister(DReg, TRI))
+ return 0;
+ }
+
+ // MI has an unwanted D-register dependency.
+ // Avoid defs in the previous N instructrions.
+ return SwiftPartialUpdateClearance;
+}
+
+// Break a partial register dependency after getPartialRegUpdateClearance
+// returned non-zero.
+void ARMBaseInstrInfo::
+breakPartialRegDependency(MachineBasicBlock::iterator MI,
+ unsigned OpNum,
+ const TargetRegisterInfo *TRI) const {
+ assert(MI && OpNum < MI->getDesc().getNumDefs() && "OpNum is not a def");
+ assert(TRI && "Need TRI instance");
+
+ const MachineOperand &MO = MI->getOperand(OpNum);
+ unsigned Reg = MO.getReg();
+ assert(TargetRegisterInfo::isPhysicalRegister(Reg) &&
+ "Can't break virtual register dependencies.");
+ unsigned DReg = Reg;
+
+ // If MI defines an S-reg, find the corresponding D super-register.
+ if (ARM::SPRRegClass.contains(Reg)) {
+ DReg = ARM::D0 + (Reg - ARM::S0) / 2;
+ assert(TRI->isSuperRegister(Reg, DReg) && "Register enums broken");
+ }