#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
+#include "llvm/ADT/BitVector.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
setOperationAction(ISD::LOAD, MVT::i1, Custom);
- setOperationAction(ISD::FP_TO_SINT, MVT::i64, Expand);
- setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
-
setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
setOperationAction(ISD::FRINT, MVT::f64, Legal);
}
- // FIXME: These should be removed and handled the same was as f32 fneg. Source
- // modifiers also work for the double instructions.
- setOperationAction(ISD::FNEG, MVT::f64, Expand);
- setOperationAction(ISD::FABS, MVT::f64, Expand);
-
setOperationAction(ISD::FDIV, MVT::f32, Custom);
+ setTargetDAGCombine(ISD::FADD);
+ setTargetDAGCombine(ISD::FSUB);
setTargetDAGCombine(ISD::SELECT_CC);
setTargetDAGCombine(ISD::SETCC);
setTargetDAGCombine(ISD::UINT_TO_FP);
+ // All memory operations. Some folding on the pointer operand is done to help
+ // matching the constant offsets in the addressing modes.
+ setTargetDAGCombine(ISD::LOAD);
+ setTargetDAGCombine(ISD::STORE);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD);
+ setTargetDAGCombine(ISD::ATOMIC_STORE);
+ setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP);
+ setTargetDAGCombine(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS);
+ setTargetDAGCombine(ISD::ATOMIC_SWAP);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_ADD);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_SUB);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_AND);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_OR);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_XOR);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
+ setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
+
setSchedulingPreference(Sched::RegPressure);
}
SDLoc SL, SDValue Chain,
unsigned Offset, bool Signed) const {
const DataLayout *DL = getDataLayout();
+ MachineFunction &MF = DAG.getMachineFunction();
+ const SIRegisterInfo *TRI =
+ static_cast<const SIRegisterInfo*>(Subtarget->getRegisterInfo());
+ unsigned InputPtrReg = TRI->getPreloadedValue(MF, SIRegisterInfo::INPUT_PTR);
Type *Ty = VT.getTypeForEVT(*DAG.getContext());
MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
PointerType *PtrTy = PointerType::get(Ty, AMDGPUAS::CONSTANT_ADDRESS);
SDValue BasePtr = DAG.getCopyFromReg(Chain, SL,
- MRI.getLiveInVirtReg(AMDGPU::SGPR0_SGPR1), MVT::i64);
+ MRI.getLiveInVirtReg(InputPtrReg), MVT::i64);
SDValue Ptr = DAG.getNode(ISD::ADD, SL, MVT::i64, BasePtr,
DAG.getConstant(Offset, MVT::i64));
SDValue PtrOffset = DAG.getUNDEF(getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
SDLoc DL, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
- const TargetRegisterInfo *TRI =
- getTargetMachine().getSubtargetImpl()->getRegisterInfo();
+ const TargetMachine &TM = getTargetMachine();
+ const SIRegisterInfo *TRI =
+ static_cast<const SIRegisterInfo*>(TM.getSubtargetImpl()->getRegisterInfo());
MachineFunction &MF = DAG.getMachineFunction();
FunctionType *FType = MF.getFunction()->getFunctionType();
assert(CallConv == CallingConv::C);
SmallVector<ISD::InputArg, 16> Splits;
- uint32_t Skipped = 0;
+ BitVector Skipped(Ins.size());
for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
const ISD::InputArg &Arg = Ins[i];
if (!Arg.Used) {
// We can savely skip PS inputs
- Skipped |= 1 << i;
+ Skipped.set(i);
++PSInputNum;
continue;
}
// The pointer to the scratch buffer is stored in SGPR2, SGPR3
if (Info->getShaderType() == ShaderType::COMPUTE) {
Info->NumUserSGPRs = 4;
- CCInfo.AllocateReg(AMDGPU::SGPR0);
- CCInfo.AllocateReg(AMDGPU::SGPR1);
- CCInfo.AllocateReg(AMDGPU::SGPR2);
- CCInfo.AllocateReg(AMDGPU::SGPR3);
- MF.addLiveIn(AMDGPU::SGPR0_SGPR1, &AMDGPU::SReg_64RegClass);
- MF.addLiveIn(AMDGPU::SGPR2_SGPR3, &AMDGPU::SReg_64RegClass);
+
+ unsigned InputPtrReg =
+ TRI->getPreloadedValue(MF, SIRegisterInfo::INPUT_PTR);
+ unsigned InputPtrRegLo =
+ TRI->getPhysRegSubReg(InputPtrReg, &AMDGPU::SReg_32RegClass, 0);
+ unsigned InputPtrRegHi =
+ TRI->getPhysRegSubReg(InputPtrReg, &AMDGPU::SReg_32RegClass, 1);
+
+ unsigned ScratchPtrReg =
+ TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
+ unsigned ScratchPtrRegLo =
+ TRI->getPhysRegSubReg(ScratchPtrReg, &AMDGPU::SReg_32RegClass, 0);
+ unsigned ScratchPtrRegHi =
+ TRI->getPhysRegSubReg(ScratchPtrReg, &AMDGPU::SReg_32RegClass, 1);
+
+ CCInfo.AllocateReg(InputPtrRegLo);
+ CCInfo.AllocateReg(InputPtrRegHi);
+ CCInfo.AllocateReg(ScratchPtrRegLo);
+ CCInfo.AllocateReg(ScratchPtrRegHi);
+ MF.addLiveIn(InputPtrReg, &AMDGPU::SReg_64RegClass);
+ MF.addLiveIn(ScratchPtrReg, &AMDGPU::SReg_64RegClass);
}
if (Info->getShaderType() == ShaderType::COMPUTE) {
for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
const ISD::InputArg &Arg = Ins[i];
- if (Skipped & (1 << i)) {
+ if (Skipped[i]) {
InVals.push_back(DAG.getUNDEF(Arg.VT));
continue;
}
if (VA.isMemLoc()) {
VT = Ins[i].VT;
EVT MemVT = Splits[i].VT;
+ const unsigned Offset = 36 + VA.getLocMemOffset();
// The first 36 bytes of the input buffer contains information about
// thread group and global sizes.
SDValue Arg = LowerParameter(DAG, VT, MemVT, DL, DAG.getRoot(),
- 36 + VA.getLocMemOffset(),
- Ins[i].Flags.isSExt());
+ Offset, Ins[i].Flags.isSExt());
+
+ const PointerType *ParamTy =
+ dyn_cast<PointerType>(FType->getParamType(Ins[i].OrigArgIndex));
+ if (Subtarget->getGeneration() == AMDGPUSubtarget::SOUTHERN_ISLANDS &&
+ ParamTy && ParamTy->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS) {
+ // On SI local pointers are just offsets into LDS, so they are always
+ // less than 16-bits. On CI and newer they could potentially be
+ // real pointers, so we can't guarantee their size.
+ Arg = DAG.getNode(ISD::AssertZext, DL, Arg.getValueType(), Arg,
+ DAG.getValueType(MVT::i16));
+ }
+
InVals.push_back(Arg);
+ Info->ABIArgOffset = Offset + MemVT.getStoreSize();
continue;
}
assert(VA.isRegLoc() && "Parameter must be in a register!");
MI->eraseFromParent();
break;
}
- case AMDGPU::FABS_SI: {
- MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
- const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
- getTargetMachine().getSubtargetImpl()->getInstrInfo());
- DebugLoc DL = MI->getDebugLoc();
- unsigned DestReg = MI->getOperand(0).getReg();
- unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
-
- BuildMI(*BB, I, DL, TII->get(AMDGPU::V_MOV_B32_e32), Reg)
- .addImm(0x7fffffff);
- BuildMI(*BB, I, DL, TII->get(AMDGPU::V_AND_B32_e32), DestReg)
- .addReg(MI->getOperand(1).getReg())
- .addReg(Reg);
- MI->eraseFromParent();
- break;
- }
- case AMDGPU::FNEG_SI: {
- MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
- const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
- getTargetMachine().getSubtargetImpl()->getInstrInfo());
- DebugLoc DL = MI->getDebugLoc();
- unsigned DestReg = MI->getOperand(0).getReg();
- unsigned Reg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
-
- BuildMI(*BB, I, DL, TII->get(AMDGPU::V_MOV_B32_e32), Reg)
- .addImm(0x80000000);
- BuildMI(*BB, I, DL, TII->get(AMDGPU::V_XOR_B32_e32), DestReg)
- .addReg(MI->getOperand(1).getReg())
- .addReg(Reg);
- MI->eraseFromParent();
- break;
- }
case AMDGPU::FCLAMP_SI: {
const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
getTargetMachine().getSubtargetImpl()->getInstrInfo());
SDValue SITargetLowering::LowerFrameIndex(SDValue Op, SelectionDAG &DAG) const {
- MachineFunction &MF = DAG.getMachineFunction();
- const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
- getTargetMachine().getSubtargetImpl()->getInstrInfo());
- const SIRegisterInfo &TRI = TII->getRegisterInfo();
FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Op);
unsigned FrameIndex = FINode->getIndex();
- CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
- TRI.getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET), MVT::i32);
-
return DAG.getTargetFrameIndex(FrameIndex, MVT::i32);
}
SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
- SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+ const SIRegisterInfo *TRI =
+ static_cast<const SIRegisterInfo*>(MF.getSubtarget().getRegisterInfo());
EVT VT = Op.getValueType();
SDLoc DL(Op);
switch (IntrinsicID) {
case Intrinsic::r600_read_ngroups_x:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 0, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::NGROUPS_X, false);
case Intrinsic::r600_read_ngroups_y:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 4, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::NGROUPS_Y, false);
case Intrinsic::r600_read_ngroups_z:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 8, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::NGROUPS_Z, false);
case Intrinsic::r600_read_global_size_x:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 12, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::GLOBAL_SIZE_X, false);
case Intrinsic::r600_read_global_size_y:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 16, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::GLOBAL_SIZE_Y, false);
case Intrinsic::r600_read_global_size_z:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 20, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::GLOBAL_SIZE_Z, false);
case Intrinsic::r600_read_local_size_x:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 24, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::LOCAL_SIZE_X, false);
case Intrinsic::r600_read_local_size_y:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 28, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::LOCAL_SIZE_Y, false);
case Intrinsic::r600_read_local_size_z:
- return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 32, false);
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ SI::KernelInputOffsets::LOCAL_SIZE_Z, false);
+
+ case Intrinsic::AMDGPU_read_workdim:
+ return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(),
+ MF.getInfo<SIMachineFunctionInfo>()->ABIArgOffset,
+ false);
+
case Intrinsic::r600_read_tgid_x:
return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
- AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 0), VT);
+ TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_X), VT);
case Intrinsic::r600_read_tgid_y:
return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
- AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 1), VT);
+ TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_Y), VT);
case Intrinsic::r600_read_tgid_z:
return CreateLiveInRegister(DAG, &AMDGPU::SReg_32RegClass,
- AMDGPU::SReg_32RegClass.getRegister(MFI->NumUserSGPRs + 2), VT);
+ TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_Z), VT);
case Intrinsic::r600_read_tidig_x:
return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
- AMDGPU::VGPR0, VT);
+ TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_X), VT);
case Intrinsic::r600_read_tidig_y:
return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
- AMDGPU::VGPR1, VT);
+ TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_Y), VT);
case Intrinsic::r600_read_tidig_z:
return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
- AMDGPU::VGPR2, VT);
+ TRI->getPreloadedValue(MF, SIRegisterInfo::TIDIG_Z), VT);
case AMDGPUIntrinsic::SI_load_const: {
SDValue Ops[] = {
Op.getOperand(1),
return SDValue();
}
+// (shl (add x, c1), c2) -> add (shl x, c2), (shl c1, c2)
+
+// This is a variant of
+// (mul (add x, c1), c2) -> add (mul x, c2), (mul c1, c2),
+//
+// The normal DAG combiner will do this, but only if the add has one use since
+// that would increase the number of instructions.
+//
+// This prevents us from seeing a constant offset that can be folded into a
+// memory instruction's addressing mode. If we know the resulting add offset of
+// a pointer can be folded into an addressing offset, we can replace the pointer
+// operand with the add of new constant offset. This eliminates one of the uses,
+// and may allow the remaining use to also be simplified.
+//
+SDValue SITargetLowering::performSHLPtrCombine(SDNode *N,
+ unsigned AddrSpace,
+ DAGCombinerInfo &DCI) const {
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+
+ if (N0.getOpcode() != ISD::ADD)
+ return SDValue();
+
+ const ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N1);
+ if (!CN1)
+ return SDValue();
+
+ const ConstantSDNode *CAdd = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+ if (!CAdd)
+ return SDValue();
+
+ const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
+ getTargetMachine().getSubtargetImpl()->getInstrInfo());
+
+ // If the resulting offset is too large, we can't fold it into the addressing
+ // mode offset.
+ APInt Offset = CAdd->getAPIntValue() << CN1->getAPIntValue();
+ if (!TII->canFoldOffset(Offset.getZExtValue(), AddrSpace))
+ return SDValue();
+
+ SelectionDAG &DAG = DCI.DAG;
+ SDLoc SL(N);
+ EVT VT = N->getValueType(0);
+
+ SDValue ShlX = DAG.getNode(ISD::SHL, SL, VT, N0.getOperand(0), N1);
+ SDValue COffset = DAG.getConstant(Offset, MVT::i32);
+
+ return DAG.getNode(ISD::ADD, SL, VT, ShlX, COffset);
+}
+
SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
case ISD::UINT_TO_FP: {
return performUCharToFloatCombine(N, DCI);
+
+ case ISD::FADD: {
+ if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
+ break;
+
+ EVT VT = N->getValueType(0);
+ if (VT != MVT::f32)
+ break;
+
+ SDValue LHS = N->getOperand(0);
+ SDValue RHS = N->getOperand(1);
+
+ // These should really be instruction patterns, but writing patterns with
+ // source modiifiers is a pain.
+
+ // fadd (fadd (a, a), b) -> mad 2.0, a, b
+ if (LHS.getOpcode() == ISD::FADD) {
+ SDValue A = LHS.getOperand(0);
+ if (A == LHS.getOperand(1)) {
+ const SDValue Two = DAG.getTargetConstantFP(2.0, MVT::f32);
+ return DAG.getNode(AMDGPUISD::MAD, DL, VT, Two, A, RHS);
+ }
+ }
+
+ // fadd (b, fadd (a, a)) -> mad 2.0, a, b
+ if (RHS.getOpcode() == ISD::FADD) {
+ SDValue A = RHS.getOperand(0);
+ if (A == RHS.getOperand(1)) {
+ const SDValue Two = DAG.getTargetConstantFP(2.0, MVT::f32);
+ return DAG.getNode(AMDGPUISD::MAD, DL, VT, Two, A, LHS);
+ }
+ }
+
+ break;
+ }
+ case ISD::FSUB: {
+ if (DCI.getDAGCombineLevel() < AfterLegalizeDAG)
+ break;
+
+ EVT VT = N->getValueType(0);
+
+ // Try to get the fneg to fold into the source modifier. This undoes generic
+ // DAG combines and folds them into the mad.
+ if (VT == MVT::f32) {
+ SDValue LHS = N->getOperand(0);
+ SDValue RHS = N->getOperand(1);
+
+ if (LHS.getOpcode() == ISD::FMUL) {
+ // (fsub (fmul a, b), c) -> mad a, b, (fneg c)
+
+ SDValue A = LHS.getOperand(0);
+ SDValue B = LHS.getOperand(1);
+ SDValue C = DAG.getNode(ISD::FNEG, DL, VT, RHS);
+
+ return DAG.getNode(AMDGPUISD::MAD, DL, VT, A, B, C);
+ }
+
+ if (RHS.getOpcode() == ISD::FMUL) {
+ // (fsub c, (fmul a, b)) -> mad (fneg a), b, c
+
+ SDValue A = DAG.getNode(ISD::FNEG, DL, VT, RHS.getOperand(0));
+ SDValue B = RHS.getOperand(1);
+ SDValue C = LHS;
+
+ return DAG.getNode(AMDGPUISD::MAD, DL, VT, A, B, C);
+ }
+
+ if (LHS.getOpcode() == ISD::FADD) {
+ // (fsub (fadd a, a), c) -> mad 2.0, a, (fneg c)
+
+ SDValue A = LHS.getOperand(0);
+ if (A == LHS.getOperand(1)) {
+ const SDValue Two = DAG.getTargetConstantFP(2.0, MVT::f32);
+ SDValue NegRHS = DAG.getNode(ISD::FNEG, DL, VT, RHS);
+
+ return DAG.getNode(AMDGPUISD::MAD, DL, VT, Two, A, NegRHS);
+ }
+ }
+
+ if (RHS.getOpcode() == ISD::FADD) {
+ // (fsub c, (fadd a, a)) -> mad -2.0, a, c
+
+ SDValue A = RHS.getOperand(0);
+ if (A == RHS.getOperand(1)) {
+ const SDValue NegTwo = DAG.getTargetConstantFP(-2.0, MVT::f32);
+ return DAG.getNode(AMDGPUISD::MAD, DL, VT, NegTwo, A, LHS);
+ }
+ }
+ }
+
+ break;
}
}
+ case ISD::LOAD:
+ case ISD::STORE:
+ case ISD::ATOMIC_LOAD:
+ case ISD::ATOMIC_STORE:
+ case ISD::ATOMIC_CMP_SWAP:
+ case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
+ case ISD::ATOMIC_SWAP:
+ case ISD::ATOMIC_LOAD_ADD:
+ case ISD::ATOMIC_LOAD_SUB:
+ case ISD::ATOMIC_LOAD_AND:
+ case ISD::ATOMIC_LOAD_OR:
+ case ISD::ATOMIC_LOAD_XOR:
+ case ISD::ATOMIC_LOAD_NAND:
+ case ISD::ATOMIC_LOAD_MIN:
+ case ISD::ATOMIC_LOAD_MAX:
+ case ISD::ATOMIC_LOAD_UMIN:
+ case ISD::ATOMIC_LOAD_UMAX: { // TODO: Target mem intrinsics.
+ if (DCI.isBeforeLegalize())
+ break;
+ MemSDNode *MemNode = cast<MemSDNode>(N);
+ SDValue Ptr = MemNode->getBasePtr();
+
+ // TODO: We could also do this for multiplies.
+ unsigned AS = MemNode->getAddressSpace();
+ if (Ptr.getOpcode() == ISD::SHL && AS != AMDGPUAS::PRIVATE_ADDRESS) {
+ SDValue NewPtr = performSHLPtrCombine(Ptr.getNode(), AS, DCI);
+ if (NewPtr) {
+ SmallVector<SDValue, 8> NewOps;
+ for (unsigned I = 0, E = MemNode->getNumOperands(); I != E; ++I)
+ NewOps.push_back(MemNode->getOperand(I));
+
+ NewOps[N->getOpcode() == ISD::STORE ? 2 : 1] = NewPtr;
+ return SDValue(DAG.UpdateNodeOperands(MemNode, NewOps), 0);
+ }
+ }
+ break;
+ }
+ }
return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
}
/// \brief Test if RegClass is one of the VSrc classes
static bool isVSrc(unsigned RegClass) {
- return AMDGPU::VSrc_32RegClassID == RegClass ||
- AMDGPU::VSrc_64RegClassID == RegClass;
+ switch(RegClass) {
+ default: return false;
+ case AMDGPU::VSrc_32RegClassID:
+ case AMDGPU::VCSrc_32RegClassID:
+ case AMDGPU::VSrc_64RegClassID:
+ case AMDGPU::VCSrc_64RegClassID:
+ return true;
+ }
}
/// \brief Test if RegClass is one of the SSrc classes
// If the COPY_TO_REGCLASS instruction is copying to a VSrc register
// class, then the register class for the value could be either a
// VReg or and SReg. In order to get a more accurate
- if (OpClassID == AMDGPU::VSrc_32RegClassID ||
- OpClassID == AMDGPU::VSrc_64RegClassID) {
+ if (isVSrc(OpClassID))
return getRegClassForNode(DAG, Op.getOperand(0));
- }
+
return TRI.getRegClass(OpClassID);
case AMDGPU::EXTRACT_SUBREG: {
int SubIdx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
return TRI->getRegClass(RegClass)->hasSubClassEq(RC);
}
-/// \brief Make sure that we don't exeed the number of allowed scalars
-void SITargetLowering::ensureSRegLimit(SelectionDAG &DAG, SDValue &Operand,
- unsigned RegClass,
- bool &ScalarSlotUsed) const {
-
- // First map the operands register class to a destination class
- if (RegClass == AMDGPU::VSrc_32RegClassID)
- RegClass = AMDGPU::VReg_32RegClassID;
- else if (RegClass == AMDGPU::VSrc_64RegClassID)
- RegClass = AMDGPU::VReg_64RegClassID;
- else
- return;
-
- // Nothing to do if they fit naturally
- if (fitsRegClass(DAG, Operand, RegClass))
- return;
-
- // If the scalar slot isn't used yet use it now
- if (!ScalarSlotUsed) {
- ScalarSlotUsed = true;
- return;
- }
-
- // This is a conservative aproach. It is possible that we can't determine the
- // correct register class and copy too often, but better safe than sorry.
-
- SDNode *Node;
- // We can't use COPY_TO_REGCLASS with FrameIndex arguments.
- if (isa<FrameIndexSDNode>(Operand)) {
- unsigned Opcode = Operand.getValueType() == MVT::i32 ?
- AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
- Node = DAG.getMachineNode(Opcode, SDLoc(), Operand.getValueType(),
- Operand);
- } else {
- SDValue RC = DAG.getTargetConstant(RegClass, MVT::i32);
- Node = DAG.getMachineNode(TargetOpcode::COPY_TO_REGCLASS, SDLoc(),
- Operand.getValueType(), Operand, RC);
- }
- Operand = SDValue(Node, 0);
-}
-
/// \returns true if \p Node's operands are different from the SDValue list
/// \p Ops
static bool isNodeChanged(const SDNode *Node, const std::vector<SDValue> &Ops) {
return false;
}
-/// \brief Try to fold the Nodes operands into the Node
-SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
- SelectionDAG &DAG) const {
-
+/// TODO: This needs to be removed. It's current primary purpose is to fold
+/// immediates into operands when legal. The legalization parts are redundant
+/// with SIInstrInfo::legalizeOperands which is called in a post-isel hook.
+SDNode *SITargetLowering::legalizeOperands(MachineSDNode *Node,
+ SelectionDAG &DAG) const {
// Original encoding (either e32 or e64)
int Opcode = Node->getMachineOpcode();
const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
assert(!DescRev || DescRev->getNumDefs() == NumDefs);
assert(!DescRev || DescRev->getNumOperands() == NumOps);
- // e64 version if available, -1 otherwise
- int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
- const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
- int InputModifiers[3] = {0};
-
- assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
-
int32_t Immediate = Desc->getSize() == 4 ? 0 : -1;
bool HaveVSrc = false, HaveSSrc = false;
// No scalar allowed when we have both VSrc and SSrc
bool ScalarSlotUsed = HaveVSrc && HaveSSrc;
+ // If this instruction has an implicit use of VCC, then it can't use the
+ // constant bus.
+ for (unsigned i = 0, e = Desc->getNumImplicitUses(); i != e; ++i) {
+ if (Desc->ImplicitUses[i] == AMDGPU::VCC) {
+ ScalarSlotUsed = true;
+ break;
+ }
+ }
+
// Second go over the operands and try to fold them
std::vector<SDValue> Ops;
- bool Promote2e64 = false;
for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
i != e && Op < NumOps; ++i, ++Op) {
// Is this a VSrc or SSrc operand?
unsigned RegClass = Desc->OpInfo[Op].RegClass;
if (isVSrc(RegClass) || isSSrc(RegClass)) {
- // Try to fold the immediates
- if (!foldImm(Ops[i], Immediate, ScalarSlotUsed)) {
- // Folding didn't work, make sure we don't hit the SReg limit.
- ensureSRegLimit(DAG, Ops[i], RegClass, ScalarSlotUsed);
- }
+ // Try to fold the immediates. If this ends up with multiple constant bus
+ // uses, it will be legalized later.
+ foldImm(Ops[i], Immediate, ScalarSlotUsed);
continue;
- } else {
- // If it's not a VSrc or SSrc operand check if we have a GlobalAddress.
- // These will be lowered to immediates, so we will need to insert a MOV.
- if (isa<GlobalAddressSDNode>(Ops[i])) {
- SDNode *Node = DAG.getMachineNode(AMDGPU::V_MOV_B32_e32, SDLoc(),
- Operand.getValueType(), Operand);
- Ops[i] = SDValue(Node, 0);
- }
}
if (i == 1 && DescRev && fitsRegClass(DAG, Ops[0], RegClass)) {
continue;
}
}
-
- if (Immediate)
- continue;
-
- if (DescE64) {
- // Test if it makes sense to switch to e64 encoding
- unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
- if (!isVSrc(OtherRegClass) && !isSSrc(OtherRegClass))
- continue;
-
- int32_t TmpImm = -1;
- if (foldImm(Ops[i], TmpImm, ScalarSlotUsed) ||
- (!fitsRegClass(DAG, Ops[i], RegClass) &&
- fitsRegClass(DAG, Ops[1], OtherRegClass))) {
-
- // Switch to e64 encoding
- Immediate = -1;
- Promote2e64 = true;
- Desc = DescE64;
- DescE64 = nullptr;
- }
- }
-
- if (!DescE64 && !Promote2e64)
- continue;
- if (!Operand.isMachineOpcode())
- continue;
- }
-
- if (Promote2e64) {
- std::vector<SDValue> OldOps(Ops);
- Ops.clear();
- bool HasModifiers = TII->hasModifiers(Desc->Opcode);
- for (unsigned i = 0; i < OldOps.size(); ++i) {
- // src_modifier
- if (HasModifiers)
- Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
- Ops.push_back(OldOps[i]);
- }
- // Add the modifier flags while promoting
- if (HasModifiers) {
- for (unsigned i = 0; i < 2; ++i)
- Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
- }
}
// Add optional chain and glue
}
}
+/// \brief Legalize target independent instructions (e.g. INSERT_SUBREG)
+/// with frame index operands.
+/// LLVM assumes that inputs are to these instructions are registers.
+void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
+ SelectionDAG &DAG) const {
+
+ SmallVector<SDValue, 8> Ops;
+ for (unsigned i = 0; i < Node->getNumOperands(); ++i) {
+ if (!isa<FrameIndexSDNode>(Node->getOperand(i))) {
+ Ops.push_back(Node->getOperand(i));
+ continue;
+ }
+
+ SDLoc DL(Node);
+ Ops.push_back(SDValue(DAG.getMachineNode(AMDGPU::S_MOV_B32, DL,
+ Node->getOperand(i).getValueType(),
+ Node->getOperand(i)), 0));
+ }
+
+ DAG.UpdateNodeOperands(Node, Ops);
+}
+
/// \brief Fold the instructions after selecting them.
SDNode *SITargetLowering::PostISelFolding(MachineSDNode *Node,
SelectionDAG &DAG) const {
if (TII->isMIMG(Node->getMachineOpcode()))
adjustWritemask(Node, DAG);
- return foldOperands(Node, DAG);
+ if (Node->getMachineOpcode() == AMDGPU::INSERT_SUBREG) {
+ legalizeTargetIndependentNode(Node, DAG);
+ return Node;
+ }
+
+ return legalizeOperands(Node, DAG);
}
/// \brief Assign the register class depending on the number of
SDNode *Node) const {
const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
getTargetMachine().getSubtargetImpl()->getInstrInfo());
- if (!TII->isMIMG(MI->getOpcode()))
- return;
- unsigned VReg = MI->getOperand(0).getReg();
- unsigned Writemask = MI->getOperand(1).getImm();
- unsigned BitsSet = 0;
- for (unsigned i = 0; i < 4; ++i)
- BitsSet += Writemask & (1 << i) ? 1 : 0;
+ TII->legalizeOperands(MI);
+
+ if (TII->isMIMG(MI->getOpcode())) {
+ unsigned VReg = MI->getOperand(0).getReg();
+ unsigned Writemask = MI->getOperand(1).getImm();
+ unsigned BitsSet = 0;
+ for (unsigned i = 0; i < 4; ++i)
+ BitsSet += Writemask & (1 << i) ? 1 : 0;
+
+ const TargetRegisterClass *RC;
+ switch (BitsSet) {
+ default: return;
+ case 1: RC = &AMDGPU::VReg_32RegClass; break;
+ case 2: RC = &AMDGPU::VReg_64RegClass; break;
+ case 3: RC = &AMDGPU::VReg_96RegClass; break;
+ }
- const TargetRegisterClass *RC;
- switch (BitsSet) {
- default: return;
- case 1: RC = &AMDGPU::VReg_32RegClass; break;
- case 2: RC = &AMDGPU::VReg_64RegClass; break;
- case 3: RC = &AMDGPU::VReg_96RegClass; break;
+ unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
+ MI->setDesc(TII->get(NewOpcode));
+ MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
+ MRI.setRegClass(VReg, RC);
+ return;
}
- unsigned NewOpcode = TII->getMaskedMIMGOp(MI->getOpcode(), BitsSet);
- MI->setDesc(TII->get(NewOpcode));
- MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
- MRI.setRegClass(VReg, RC);
+ // Replace unused atomics with the no return version.
+ int NoRetAtomicOp = AMDGPU::getAtomicNoRetOp(MI->getOpcode());
+ if (NoRetAtomicOp != -1) {
+ if (!Node->hasAnyUseOfValue(0)) {
+ MI->setDesc(TII->get(NoRetAtomicOp));
+ MI->RemoveOperand(0);
+ }
+
+ return;
+ }
}
MachineSDNode *SITargetLowering::AdjustRegClass(MachineSDNode *N,
return N;
}
ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
- SDValue Ops[] = {
- SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
- DAG.getConstant(0, MVT::i64)), 0),
- N->getOperand(0),
- DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
- };
+ MachineSDNode *RSrc = DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL,
+ MVT::i128,
+ DAG.getConstant(0, MVT::i64));
+
+ SmallVector<SDValue, 8> Ops;
+ Ops.push_back(SDValue(RSrc, 0));
+ Ops.push_back(N->getOperand(0));
+ Ops.push_back(DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32));
+
+ // Copy remaining operands so we keep any chain and glue nodes that follow
+ // the normal operands.
+ for (unsigned I = 2, E = N->getNumOperands(); I != E; ++I)
+ Ops.push_back(N->getOperand(I));
+
return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
}
}