setOperationAction(ISD::SETCC, MVT::v2i1, Expand);
setOperationAction(ISD::SETCC, MVT::v4i1, Expand);
+ setOperationAction(ISD::BSWAP, MVT::i32, Legal);
+
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Legal);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
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::FDIV, MVT::f32, Custom);
+ setTargetDAGCombine(ISD::FADD);
setTargetDAGCombine(ISD::FSUB);
setTargetDAGCombine(ISD::SELECT_CC);
setTargetDAGCombine(ISD::SETCC);
// TargetLowering queries
//===----------------------------------------------------------------------===//
+bool SITargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &,
+ EVT) const {
+ // SI has some legal vector types, but no legal vector operations. Say no
+ // shuffles are legal in order to prefer scalarizing some vector operations.
+ return false;
+}
+
// FIXME: This really needs an address space argument. The immediate offset
// size is different for different sets of memory instruction sets.
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));
}
InVals.push_back(Arg);
+ Info->ABIArgOffset = Offset + MemVT.getStoreSize();
continue;
}
assert(VA.isRegLoc() && "Parameter must be in a register!");
MI->eraseFromParent();
break;
}
- case AMDGPU::SI_BUFFER_RSRC: {
- unsigned SuperReg = MI->getOperand(0).getReg();
- unsigned Args[4];
- for (unsigned i = 0, e = 4; i < e; ++i) {
- MachineOperand &Arg = MI->getOperand(i + 1);
-
- if (Arg.isReg()) {
- Args[i] = Arg.getReg();
- continue;
- }
-
- assert(Arg.isImm());
- unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
- BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::S_MOV_B32), Reg)
- .addImm(Arg.getImm());
- Args[i] = Reg;
- }
- BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::REG_SEQUENCE),
- SuperReg)
- .addReg(Args[0])
- .addImm(AMDGPU::sub0)
- .addReg(Args[1])
- .addImm(AMDGPU::sub1)
- .addReg(Args[2])
- .addImm(AMDGPU::sub2)
- .addReg(Args[3])
- .addImm(AMDGPU::sub3);
- MI->eraseFromParent();
- break;
- }
case AMDGPU::V_SUB_F64: {
unsigned DestReg = MI->getOperand(0).getReg();
BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::V_ADD_F64), DestReg)
case Intrinsic::r600_read_local_size_z:
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,
TRI->getPreloadedValue(MF, SIRegisterInfo::TGID_X), VT);
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;
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;
/// \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) ||
- isa<GlobalAddressSDNode>(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 commute instructions and insert copies in order to satisfy the
-/// operand constraints.
+/// 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)
// 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;
for (unsigned i = 0, e = Node->getNumOperands(), Op = NumDefs;
// 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)) {
}
}
+/// \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);
+ if (Node->getMachineOpcode() == AMDGPU::INSERT_SUBREG) {
+ legalizeTargetIndependentNode(Node, DAG);
+ return Node;
+ }
+
return legalizeOperands(Node, DAG);
}
const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
getTargetMachine().getSubtargetImpl()->getInstrInfo());
+ TII->legalizeOperands(MI);
+
if (TII->isMIMG(MI->getOpcode())) {
unsigned VReg = MI->getOperand(0).getReg();
unsigned Writemask = MI->getOperand(1).getImm();
projects / oota-llvm.git / blobdiff