//
//===----------------------------------------------------------------------===//
//
-// Most of the DAG lowering is handled in AMDILISelLowering.cpp. This file
+// Most of the DAG lowering is handled in AMDGPUISelLowering.cpp. This file
// is mostly EmitInstrWithCustomInserter().
//
//===----------------------------------------------------------------------===//
TII(static_cast<const R600InstrInfo*>(TM.getInstrInfo()))
{
setOperationAction(ISD::MUL, MVT::i64, Expand);
-// setSchedulingPreference(Sched::VLIW);
- addRegisterClass(MVT::v4f32, &AMDIL::R600_Reg128RegClass);
- addRegisterClass(MVT::f32, &AMDIL::R600_Reg32RegClass);
- addRegisterClass(MVT::v4i32, &AMDIL::R600_Reg128RegClass);
- addRegisterClass(MVT::i32, &AMDIL::R600_Reg32RegClass);
-
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Legal);
- setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Legal);
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Legal);
- setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Legal);
+ addRegisterClass(MVT::v4f32, &AMDGPU::R600_Reg128RegClass);
+ addRegisterClass(MVT::f32, &AMDGPU::R600_Reg32RegClass);
+ addRegisterClass(MVT::v4i32, &AMDGPU::R600_Reg128RegClass);
+ addRegisterClass(MVT::i32, &AMDGPU::R600_Reg32RegClass);
+ computeRegisterProperties();
setOperationAction(ISD::FSUB, MVT::f32, Expand);
-#if 0
+ setOperationAction(ISD::ROTL, MVT::i32, Custom);
- setTargetDAGCombine(ISD::Constant);
- setTargetDAGCombine(ISD::ConstantFP);
+ setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
+ setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
-#endif
+ setOperationAction(ISD::SETCC, MVT::i32, Custom);
+
+ setSchedulingPreference(Sched::VLIW);
}
MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
switch (MI->getOpcode()) {
default: return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
- case AMDIL::TGID_X:
- addLiveIn(MI, MF, MRI, TII, AMDIL::T1_X);
+ case AMDGPU::TGID_X:
+ addLiveIn(MI, MF, MRI, TII, AMDGPU::T1_X);
break;
- case AMDIL::TGID_Y:
- addLiveIn(MI, MF, MRI, TII, AMDIL::T1_Y);
+ case AMDGPU::TGID_Y:
+ addLiveIn(MI, MF, MRI, TII, AMDGPU::T1_Y);
break;
- case AMDIL::TGID_Z:
- addLiveIn(MI, MF, MRI, TII, AMDIL::T1_Z);
+ case AMDGPU::TGID_Z:
+ addLiveIn(MI, MF, MRI, TII, AMDGPU::T1_Z);
break;
- case AMDIL::TIDIG_X:
- addLiveIn(MI, MF, MRI, TII, AMDIL::T0_X);
+ case AMDGPU::TIDIG_X:
+ addLiveIn(MI, MF, MRI, TII, AMDGPU::T0_X);
break;
- case AMDIL::TIDIG_Y:
- addLiveIn(MI, MF, MRI, TII, AMDIL::T0_Y);
+ case AMDGPU::TIDIG_Y:
+ addLiveIn(MI, MF, MRI, TII, AMDGPU::T0_Y);
break;
- case AMDIL::TIDIG_Z:
- addLiveIn(MI, MF, MRI, TII, AMDIL::T0_Z);
+ case AMDGPU::TIDIG_Z:
+ addLiveIn(MI, MF, MRI, TII, AMDGPU::T0_Z);
break;
- case AMDIL::NGROUPS_X:
+ case AMDGPU::NGROUPS_X:
lowerImplicitParameter(MI, *BB, MRI, 0);
break;
- case AMDIL::NGROUPS_Y:
+ case AMDGPU::NGROUPS_Y:
lowerImplicitParameter(MI, *BB, MRI, 1);
break;
- case AMDIL::NGROUPS_Z:
+ case AMDGPU::NGROUPS_Z:
lowerImplicitParameter(MI, *BB, MRI, 2);
break;
- case AMDIL::GLOBAL_SIZE_X:
+ case AMDGPU::GLOBAL_SIZE_X:
lowerImplicitParameter(MI, *BB, MRI, 3);
break;
- case AMDIL::GLOBAL_SIZE_Y:
+ case AMDGPU::GLOBAL_SIZE_Y:
lowerImplicitParameter(MI, *BB, MRI, 4);
break;
- case AMDIL::GLOBAL_SIZE_Z:
+ case AMDGPU::GLOBAL_SIZE_Z:
lowerImplicitParameter(MI, *BB, MRI, 5);
break;
- case AMDIL::LOCAL_SIZE_X:
+ case AMDGPU::LOCAL_SIZE_X:
lowerImplicitParameter(MI, *BB, MRI, 6);
break;
- case AMDIL::LOCAL_SIZE_Y:
+ case AMDGPU::LOCAL_SIZE_Y:
lowerImplicitParameter(MI, *BB, MRI, 7);
break;
- case AMDIL::LOCAL_SIZE_Z:
+ case AMDGPU::LOCAL_SIZE_Z:
lowerImplicitParameter(MI, *BB, MRI, 8);
break;
- case AMDIL::CLAMP_R600:
+ case AMDGPU::CLAMP_R600:
MI->getOperand(0).addTargetFlag(MO_FLAG_CLAMP);
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::MOV))
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::MOV))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1));
break;
- case AMDIL::FABS_R600:
+ case AMDGPU::FABS_R600:
MI->getOperand(1).addTargetFlag(MO_FLAG_ABS);
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::MOV))
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::MOV))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1));
break;
- case AMDIL::FNEG_R600:
+ case AMDGPU::FNEG_R600:
MI->getOperand(1).addTargetFlag(MO_FLAG_NEG);
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::MOV))
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::MOV))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1));
break;
- case AMDIL::R600_LOAD_CONST:
+ case AMDGPU::R600_LOAD_CONST:
{
int64_t RegIndex = MI->getOperand(1).getImm();
- unsigned ConstantReg = AMDIL::R600_CReg32RegClass.getRegister(RegIndex);
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::COPY))
+ unsigned ConstantReg = AMDGPU::R600_CReg32RegClass.getRegister(RegIndex);
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::COPY))
.addOperand(MI->getOperand(0))
.addReg(ConstantReg);
break;
}
- case AMDIL::LOAD_INPUT:
+ case AMDGPU::LOAD_INPUT:
{
int64_t RegIndex = MI->getOperand(1).getImm();
addLiveIn(MI, MF, MRI, TII,
- AMDIL::R600_TReg32RegClass.getRegister(RegIndex));
+ AMDGPU::R600_TReg32RegClass.getRegister(RegIndex));
break;
}
- case AMDIL::MASK_WRITE:
+ case AMDGPU::MASK_WRITE:
{
unsigned maskedRegister = MI->getOperand(0).getReg();
assert(TargetRegisterInfo::isVirtualRegister(maskedRegister));
return BB;
}
- case AMDIL::STORE_OUTPUT:
+ case AMDGPU::RAT_WRITE_CACHELESS_eg:
+ {
+ // Convert to DWORD address
+ unsigned NewAddr = MRI.createVirtualRegister(
+ AMDGPU::R600_TReg32_XRegisterClass);
+ unsigned ShiftValue = MRI.createVirtualRegister(
+ AMDGPU::R600_TReg32RegisterClass);
+
+ // XXX In theory, we should be able to pass ShiftValue directly to
+ // the LSHR_eg instruction as an inline literal, but I tried doing it
+ // this way and it didn't produce the correct results.
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::MOV), ShiftValue)
+ .addReg(AMDGPU::ALU_LITERAL_X)
+ .addImm(2);
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::LSHR_eg), NewAddr)
+ .addOperand(MI->getOperand(1))
+ .addReg(ShiftValue);
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
+ .addOperand(MI->getOperand(0))
+ .addReg(NewAddr);
+ break;
+ }
+
+ case AMDGPU::STORE_OUTPUT:
{
int64_t OutputIndex = MI->getOperand(1).getImm();
- unsigned OutputReg = AMDIL::R600_TReg32RegClass.getRegister(OutputIndex);
+ unsigned OutputReg = AMDGPU::R600_TReg32RegClass.getRegister(OutputIndex);
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::COPY), OutputReg)
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::COPY), OutputReg)
.addOperand(MI->getOperand(0));
if (!MRI.isLiveOut(OutputReg)) {
break;
}
- case AMDIL::RESERVE_REG:
+ case AMDGPU::RESERVE_REG:
{
R600MachineFunctionInfo * MFI = MF->getInfo<R600MachineFunctionInfo>();
int64_t ReservedIndex = MI->getOperand(0).getImm();
unsigned ReservedReg =
- AMDIL::R600_TReg32RegClass.getRegister(ReservedIndex);
+ AMDGPU::R600_TReg32RegClass.getRegister(ReservedIndex);
MFI->ReservedRegs.push_back(ReservedReg);
break;
}
- case AMDIL::TXD:
+ case AMDGPU::TXD:
{
- unsigned t0 = MRI.createVirtualRegister(AMDIL::R600_Reg128RegisterClass);
- unsigned t1 = MRI.createVirtualRegister(AMDIL::R600_Reg128RegisterClass);
+ unsigned t0 = MRI.createVirtualRegister(AMDGPU::R600_Reg128RegisterClass);
+ unsigned t1 = MRI.createVirtualRegister(AMDGPU::R600_Reg128RegisterClass);
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::TEX_SET_GRADIENTS_H), t0)
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), t0)
.addOperand(MI->getOperand(3))
.addOperand(MI->getOperand(4))
.addOperand(MI->getOperand(5));
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::TEX_SET_GRADIENTS_V), t1)
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), t1)
.addOperand(MI->getOperand(2))
.addOperand(MI->getOperand(4))
.addOperand(MI->getOperand(5));
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::TEX_SAMPLE_G))
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_G))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1))
.addOperand(MI->getOperand(4))
.addReg(t1, RegState::Implicit);
break;
}
- case AMDIL::TXD_SHADOW:
+ case AMDGPU::TXD_SHADOW:
{
- unsigned t0 = MRI.createVirtualRegister(AMDIL::R600_Reg128RegisterClass);
- unsigned t1 = MRI.createVirtualRegister(AMDIL::R600_Reg128RegisterClass);
+ unsigned t0 = MRI.createVirtualRegister(AMDGPU::R600_Reg128RegisterClass);
+ unsigned t1 = MRI.createVirtualRegister(AMDGPU::R600_Reg128RegisterClass);
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::TEX_SET_GRADIENTS_H), t0)
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), t0)
.addOperand(MI->getOperand(3))
.addOperand(MI->getOperand(4))
.addOperand(MI->getOperand(5));
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::TEX_SET_GRADIENTS_V), t1)
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), t1)
.addOperand(MI->getOperand(2))
.addOperand(MI->getOperand(4))
.addOperand(MI->getOperand(5));
- BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::TEX_SAMPLE_C_G))
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_C_G))
.addOperand(MI->getOperand(0))
.addOperand(MI->getOperand(1))
.addOperand(MI->getOperand(4))
MachineRegisterInfo & MRI, unsigned dword_offset) const
{
MachineBasicBlock::iterator I = *MI;
- unsigned offsetReg = MRI.createVirtualRegister(&AMDIL::R600_TReg32_XRegClass);
- MRI.setRegClass(MI->getOperand(0).getReg(), &AMDIL::R600_TReg32_XRegClass);
+ unsigned PtrReg = MRI.createVirtualRegister(&AMDGPU::R600_TReg32_XRegClass);
+ MRI.setRegClass(MI->getOperand(0).getReg(), &AMDGPU::R600_TReg32_XRegClass);
- BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDIL::MOV), offsetReg)
- .addReg(AMDIL::ALU_LITERAL_X)
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::MOV), PtrReg)
+ .addReg(AMDGPU::ALU_LITERAL_X)
.addImm(dword_offset * 4);
- BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDIL::VTX_READ_eg))
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::VTX_READ_PARAM_i32_eg))
.addOperand(MI->getOperand(0))
- .addReg(offsetReg)
+ .addReg(PtrReg)
.addImm(0);
}
+
+//===----------------------------------------------------------------------===//
+// Custom DAG Lowering Operations
+//===----------------------------------------------------------------------===//
+
+
+SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const
+{
+ switch (Op.getOpcode()) {
+ default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
+ case ISD::ROTL: return LowerROTL(Op, DAG);
+ case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
+ case ISD::SETCC: return LowerSETCC(Op, DAG);
+ }
+}
+
+SDValue R600TargetLowering::LowerROTL(SDValue Op, SelectionDAG &DAG) const
+{
+ DebugLoc DL = Op.getDebugLoc();
+ EVT VT = Op.getValueType();
+
+ return DAG.getNode(AMDGPUISD::BITALIGN, DL, VT,
+ Op.getOperand(0),
+ Op.getOperand(0),
+ DAG.getNode(ISD::SUB, DL, VT,
+ DAG.getConstant(32, MVT::i32),
+ Op.getOperand(1)));
+}
+
+SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
+{
+ DebugLoc DL = Op.getDebugLoc();
+ EVT VT = Op.getValueType();
+
+ SDValue LHS = Op.getOperand(0);
+ SDValue RHS = Op.getOperand(1);
+ SDValue True = Op.getOperand(2);
+ SDValue False = Op.getOperand(3);
+ SDValue CC = Op.getOperand(4);
+ ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
+ SDValue Temp;
+
+ // LHS and RHS are guaranteed to be the same value type
+ EVT CompareVT = LHS.getValueType();
+
+ // We need all the operands of SELECT_CC to have the same value type, so if
+ // necessary we need to convert LHS and RHS to be the same type True and
+ // False. True and False are guaranteed to have the same type as this
+ // SELECT_CC node.
+
+ if (CompareVT != VT) {
+ ISD::NodeType ConversionOp = ISD::DELETED_NODE;
+ if (VT == MVT::f32 && CompareVT == MVT::i32) {
+ if (isUnsignedIntSetCC(CCOpcode)) {
+ ConversionOp = ISD::UINT_TO_FP;
+ } else {
+ ConversionOp = ISD::SINT_TO_FP;
+ }
+ } else if (VT == MVT::i32 && CompareVT == MVT::f32) {
+ ConversionOp = ISD::FP_TO_SINT;
+ } else {
+ // I don't think there will be any other type pairings.
+ assert(!"Unhandled operand type parings in SELECT_CC");
+ }
+ // XXX Check the value of LHS and RHS and avoid creating sequences like
+ // (FTOI (ITOF))
+ LHS = DAG.getNode(ConversionOp, DL, VT, LHS);
+ RHS = DAG.getNode(ConversionOp, DL, VT, RHS);
+ }
+
+ // If True is a hardware TRUE value and False is a hardware FALSE value or
+ // vice-versa we can handle this with a native instruction (SET* instructions).
+ if ((isHWTrueValue(True) && isHWFalseValue(False))) {
+ return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
+ }
+
+ // XXX If True is a hardware TRUE value and False is a hardware FALSE value,
+ // we can handle this with a native instruction, but we need to swap true
+ // and false and change the conditional.
+ if (isHWTrueValue(False) && isHWFalseValue(True)) {
+ }
+
+ // XXX Check if we can lower this to a SELECT or if it is supported by a native
+ // operation. (The code below does this but we don't have the Instruction
+ // selection patterns to do this yet.
+#if 0
+ if (isZero(LHS) || isZero(RHS)) {
+ SDValue Cond = (isZero(LHS) ? RHS : LHS);
+ bool SwapTF = false;
+ switch (CCOpcode) {
+ case ISD::SETOEQ:
+ case ISD::SETUEQ:
+ case ISD::SETEQ:
+ SwapTF = true;
+ // Fall through
+ case ISD::SETONE:
+ case ISD::SETUNE:
+ case ISD::SETNE:
+ // We can lower to select
+ if (SwapTF) {
+ Temp = True;
+ True = False;
+ False = Temp;
+ }
+ // CNDE
+ return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
+ default:
+ // Supported by a native operation (CNDGE, CNDGT)
+ return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
+ }
+ }
+#endif
+
+ // If we make it this for it means we have no native instructions to handle
+ // this SELECT_CC, so we must lower it.
+ SDValue HWTrue, HWFalse;
+
+ if (VT == MVT::f32) {
+ HWTrue = DAG.getConstantFP(1.0f, VT);
+ HWFalse = DAG.getConstantFP(0.0f, VT);
+ } else if (VT == MVT::i32) {
+ HWTrue = DAG.getConstant(-1, VT);
+ HWFalse = DAG.getConstant(0, VT);
+ }
+ else {
+ assert(!"Unhandled value type in LowerSELECT_CC");
+ }
+
+ // Lower this unsupported SELECT_CC into a combination of two supported
+ // SELECT_CC operations.
+ SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, HWTrue, HWFalse, CC);
+
+ return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
+}
+
+SDValue R600TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const
+{
+ SDValue Cond;
+ SDValue LHS = Op.getOperand(0);
+ SDValue RHS = Op.getOperand(1);
+ SDValue CC = Op.getOperand(2);
+ DebugLoc DL = Op.getDebugLoc();
+ assert(Op.getValueType() == MVT::i32);
+ Cond = DAG.getNode(
+ ISD::SELECT_CC,
+ Op.getDebugLoc(),
+ MVT::i32,
+ LHS, RHS,
+ DAG.getConstant(-1, MVT::i32),
+ DAG.getConstant(0, MVT::i32),
+ CC);
+ Cond = DAG.getNode(
+ ISD::AND,
+ DL,
+ MVT::i32,
+ DAG.getConstant(1, MVT::i32),
+ Cond);
+ return Cond;
+}