//===----------------------------------------------------------------------===//
#include "SIISelLowering.h"
+#include "AMDIL.h"
+#include "AMDILIntrinsicInfo.h"
#include "SIInstrInfo.h"
+#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/SelectionDAG.h"
using namespace llvm;
setOperationAction(ISD::BR_CC, MVT::i32, Custom);
+ setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+
+ // We need to custom lower loads from the USER_SGPR address space, so we can
+ // add the SGPRs as livein registers.
+ setOperationAction(ISD::LOAD, MVT::i32, Custom);
+ setOperationAction(ISD::LOAD, MVT::i64, Custom);
+
setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
+ setTargetDAGCombine(ISD::SELECT_CC);
+
+ setTargetDAGCombine(ISD::SETCC);
}
MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
MI->eraseFromParent();
break;
+ case AMDGPU::FNEG_SI:
+ BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64))
+ .addOperand(MI->getOperand(0))
+ .addOperand(MI->getOperand(1))
+ // VSRC1-2 are unused, but we still need to fill all the
+ // operand slots, so we just reuse the VSRC0 operand
+ .addOperand(MI->getOperand(1))
+ .addOperand(MI->getOperand(1))
+ .addImm(0) // ABS
+ .addImm(0) // CLAMP
+ .addImm(0) // OMOD
+ .addImm(1); // NEG
+ MI->eraseFromParent();
+ break;
+ case AMDGPU::SHADER_TYPE:
+ BB->getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType =
+ MI->getOperand(0).getImm();
+ MI->eraseFromParent();
+ break;
+
case AMDGPU::SI_INTERP:
LowerSI_INTERP(MI, *BB, I, MRI);
break;
case AMDGPU::SI_INTERP_CONST:
- LowerSI_INTERP_CONST(MI, *BB, I);
+ LowerSI_INTERP_CONST(MI, *BB, I, MRI);
break;
- case AMDGPU::SI_V_CNDLT:
- LowerSI_V_CNDLT(MI, *BB, I, MRI);
+ case AMDGPU::SI_KIL:
+ LowerSI_KIL(MI, *BB, I, MRI);
break;
- case AMDGPU::USE_SGPR_32:
- case AMDGPU::USE_SGPR_64:
- lowerUSE_SGPR(MI, BB->getParent(), MRI);
- MI->eraseFromParent();
+ case AMDGPU::SI_WQM:
+ LowerSI_WQM(MI, *BB, I, MRI);
break;
- case AMDGPU::VS_LOAD_BUFFER_INDEX:
- addLiveIn(MI, BB->getParent(), MRI, TII, AMDGPU::VGPR0);
- MI->eraseFromParent();
+ case AMDGPU::SI_V_CNDLT:
+ LowerSI_V_CNDLT(MI, *BB, I, MRI);
break;
}
return BB;
.addImm(0);
}
+
+void SITargetLowering::LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
+ MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
+{
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WQM_B64), AMDGPU::EXEC)
+ .addReg(AMDGPU::EXEC);
+
+ MI->eraseFromParent();
+}
+
void SITargetLowering::LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
{
unsigned tmp = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
+ unsigned M0 = MRI.createVirtualRegister(&AMDGPU::M0RegRegClass);
MachineOperand dst = MI->getOperand(0);
MachineOperand iReg = MI->getOperand(1);
MachineOperand jReg = MI->getOperand(2);
MachineOperand attr = MI->getOperand(4);
MachineOperand params = MI->getOperand(5);
- BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32))
- .addReg(AMDGPU::M0)
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32), M0)
.addOperand(params);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P1_F32), tmp)
.addOperand(iReg)
.addOperand(attr_chan)
- .addOperand(attr);
+ .addOperand(attr)
+ .addReg(M0);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P2_F32))
.addOperand(dst)
.addReg(tmp)
.addOperand(jReg)
.addOperand(attr_chan)
- .addOperand(attr);
+ .addOperand(attr)
+ .addReg(M0);
MI->eraseFromParent();
}
void SITargetLowering::LowerSI_INTERP_CONST(MachineInstr *MI,
- MachineBasicBlock &BB, MachineBasicBlock::iterator I) const
+ MachineBasicBlock &BB, MachineBasicBlock::iterator I,
+ MachineRegisterInfo &MRI) const
{
MachineOperand dst = MI->getOperand(0);
MachineOperand attr_chan = MI->getOperand(1);
MachineOperand attr = MI->getOperand(2);
MachineOperand params = MI->getOperand(3);
+ unsigned M0 = MRI.createVirtualRegister(&AMDGPU::M0RegRegClass);
- BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32))
- .addReg(AMDGPU::M0)
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32), M0)
.addOperand(params);
BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_MOV_F32))
.addOperand(dst)
.addOperand(attr_chan)
- .addOperand(attr);
+ .addOperand(attr)
+ .addReg(M0);
MI->eraseFromParent();
}
-void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
+void SITargetLowering::LowerSI_KIL(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
{
- BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMP_LT_F32_e32))
- .addOperand(MI->getOperand(1))
+ // Clear this pixel from the exec mask if the operand is negative
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMPX_LE_F32_e32),
+ AMDGPU::VCC)
+ .addReg(AMDGPU::SREG_LIT_0)
+ .addOperand(MI->getOperand(0));
+
+ // If the exec mask is non-zero, skip the next two instructions
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_CBRANCH_EXECNZ))
+ .addImm(3)
+ .addReg(AMDGPU::EXEC);
+
+ // Exec mask is zero: Export to NULL target...
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::EXP))
+ .addImm(0)
+ .addImm(0x09) // V_008DFC_SQ_EXP_NULL
+ .addImm(0)
+ .addImm(1)
+ .addImm(1)
+ .addReg(AMDGPU::SREG_LIT_0)
+ .addReg(AMDGPU::SREG_LIT_0)
+ .addReg(AMDGPU::SREG_LIT_0)
.addReg(AMDGPU::SREG_LIT_0);
- BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CNDMASK_B32))
- .addOperand(MI->getOperand(0))
- .addOperand(MI->getOperand(2))
- .addOperand(MI->getOperand(3));
+ // ... and terminate wavefront
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_ENDPGM));
MI->eraseFromParent();
}
-void SITargetLowering::lowerUSE_SGPR(MachineInstr *MI,
- MachineFunction * MF, MachineRegisterInfo & MRI) const
+void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
+ MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const
{
- const TargetInstrInfo * TII = getTargetMachine().getInstrInfo();
- unsigned dstReg = MI->getOperand(0).getReg();
- int64_t newIndex = MI->getOperand(1).getImm();
- const TargetRegisterClass * dstClass = MRI.getRegClass(dstReg);
- unsigned DwordWidth = dstClass->getSize() / 4;
- assert(newIndex % DwordWidth == 0 && "USER_SGPR not properly aligned");
- newIndex = newIndex / DwordWidth;
-
- unsigned newReg = dstClass->getRegister(newIndex);
- addLiveIn(MI, MF, MRI, TII, newReg);
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMP_GT_F32_e32),
+ AMDGPU::VCC)
+ .addReg(AMDGPU::SREG_LIT_0)
+ .addOperand(MI->getOperand(1));
+
+ BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CNDMASK_B32))
+ .addOperand(MI->getOperand(0))
+ .addOperand(MI->getOperand(2))
+ .addOperand(MI->getOperand(3))
+ .addReg(AMDGPU::VCC);
+
+ MI->eraseFromParent();
}
EVT SITargetLowering::getSetCCResultType(EVT VT) const
switch (Op.getOpcode()) {
default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
case ISD::BR_CC: return LowerBR_CC(Op, DAG);
+ case ISD::LOAD: return LowerLOAD(Op, DAG);
case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
case ISD::AND: return Loweri1ContextSwitch(Op, DAG, ISD::AND);
+ case ISD::INTRINSIC_WO_CHAIN: {
+ unsigned IntrinsicID =
+ cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
+ EVT VT = Op.getValueType();
+ switch (IntrinsicID) {
+ case AMDGPUIntrinsic::SI_vs_load_buffer_index:
+ return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
+ AMDGPU::VGPR0, VT);
+ default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
+ }
+ break;
+ }
}
+ return SDValue();
}
/// Loweri1ContextSwitch - The function is for lowering i1 operations on the
CC);
Result = DAG.getNode(
- AMDILISD::BRANCH_COND,
+ AMDGPUISD::BRANCH_COND,
CmpValue.getDebugLoc(),
MVT::Other, Chain,
JumpT, CmpValue);
return Result;
}
+SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const
+{
+ EVT VT = Op.getValueType();
+ LoadSDNode *Ptr = dyn_cast<LoadSDNode>(Op);
+
+ assert(Ptr);
+
+ unsigned AddrSpace = Ptr->getPointerInfo().getAddrSpace();
+
+ // We only need to lower USER_SGPR address space loads
+ if (AddrSpace != AMDGPUAS::USER_SGPR_ADDRESS) {
+ return SDValue();
+ }
+
+ // Loads from the USER_SGPR address space can only have constant value
+ // pointers.
+ ConstantSDNode *BasePtr = dyn_cast<ConstantSDNode>(Ptr->getBasePtr());
+ assert(BasePtr);
+
+ unsigned TypeDwordWidth = VT.getSizeInBits() / 32;
+ const TargetRegisterClass * dstClass;
+ switch (TypeDwordWidth) {
+ default:
+ assert(!"USER_SGPR value size not implemented");
+ return SDValue();
+ case 1:
+ dstClass = &AMDGPU::SReg_32RegClass;
+ break;
+ case 2:
+ dstClass = &AMDGPU::SReg_64RegClass;
+ break;
+ }
+ uint64_t Index = BasePtr->getZExtValue();
+ assert(Index % TypeDwordWidth == 0 && "USER_SGPR not properly aligned");
+ unsigned SGPRIndex = Index / TypeDwordWidth;
+ unsigned Reg = dstClass->getRegister(SGPRIndex);
+
+ DAG.ReplaceAllUsesOfValueWith(Op, CreateLiveInRegister(DAG, dstClass, Reg,
+ VT));
+ return SDValue();
+}
+
SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
{
SDValue LHS = Op.getOperand(0);
return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
}
+//===----------------------------------------------------------------------===//
+// Custom DAG optimizations
+//===----------------------------------------------------------------------===//
+
+SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const {
+ SelectionDAG &DAG = DCI.DAG;
+ DebugLoc DL = N->getDebugLoc();
+ EVT VT = N->getValueType(0);
+
+ switch (N->getOpcode()) {
+ default: break;
+ case ISD::SELECT_CC: {
+ N->dump();
+ ConstantSDNode *True, *False;
+ // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
+ if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
+ && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
+ && True->isAllOnesValue()
+ && False->isNullValue()
+ && VT == MVT::i1) {
+ return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
+ N->getOperand(1), N->getOperand(4));
+
+ }
+ break;
+ }
+ case ISD::SETCC: {
+ SDValue Arg0 = N->getOperand(0);
+ SDValue Arg1 = N->getOperand(1);
+ SDValue CC = N->getOperand(2);
+ ConstantSDNode * C = NULL;
+ ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
+
+ // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
+ if (VT == MVT::i1
+ && Arg0.getOpcode() == ISD::SIGN_EXTEND
+ && Arg0.getOperand(0).getValueType() == MVT::i1
+ && (C = dyn_cast<ConstantSDNode>(Arg1))
+ && C->isNullValue()
+ && CCOp == ISD::SETNE) {
+ return SimplifySetCC(VT, Arg0.getOperand(0),
+ DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
+ }
+ break;
+ }
+ }
+ return SDValue();
+}
+
#define NODE_NAME_CASE(node) case SIISD::node: return #node;
const char* SITargetLowering::getTargetNodeName(unsigned Opcode) const
projects / mesa.git / blobdiff