setOperationAction(ISD::ADD, MVT::i64, Legal);
setOperationAction(ISD::ADD, MVT::i32, Legal);
- 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
MachineRegisterInfo & MRI = BB->getParent()->getRegInfo();
MachineBasicBlock::iterator I = MI;
- if (TII->get(MI->getOpcode()).TSFlags & SIInstrFlags::NEED_WAIT) {
- AppendS_WAITCNT(MI, *BB, llvm::next(I));
- return BB;
- }
-
switch (MI->getOpcode()) {
default:
return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
case AMDGPU::SI_INTERP_CONST:
LowerSI_INTERP_CONST(MI, *BB, I, MRI);
break;
- case AMDGPU::SI_KIL:
- LowerSI_KIL(MI, *BB, I, MRI);
- break;
case AMDGPU::SI_WQM:
LowerSI_WQM(MI, *BB, I, MRI);
break;
return BB;
}
-void SITargetLowering::AppendS_WAITCNT(MachineInstr *MI, MachineBasicBlock &BB,
- MachineBasicBlock::iterator I) const {
- BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WAITCNT))
- .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)
MI->eraseFromParent();
}
-void SITargetLowering::LowerSI_KIL(MachineInstr *MI, MachineBasicBlock &BB,
- MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
- // 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));
-
- MI->eraseFromParent();
-}
-
void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
unsigned VCC = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
switch (Op.getOpcode()) {
default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
- case ISD::BR_CC: return LowerBR_CC(Op, DAG);
+ case ISD::BRCOND: return LowerBRCOND(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);
return DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i1, OpNode);
}
-SDValue SITargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
- SDValue Chain = Op.getOperand(0);
- SDValue CC = Op.getOperand(1);
- SDValue LHS = Op.getOperand(2);
- SDValue RHS = Op.getOperand(3);
- SDValue JumpT = Op.getOperand(4);
- SDValue CmpValue;
- SDValue Result;
- CmpValue = DAG.getNode(
- ISD::SETCC,
- Op.getDebugLoc(),
- MVT::i1,
- LHS, RHS,
- CC);
-
- Result = DAG.getNode(
- AMDGPUISD::BRANCH_COND,
- CmpValue.getDebugLoc(),
- MVT::Other, Chain,
- JumpT, CmpValue);
- return Result;
+/// \brief Helper function for LowerBRCOND
+static SDNode *findUser(SDValue Value, unsigned Opcode) {
+
+ SDNode *Parent = Value.getNode();
+ for (SDNode::use_iterator I = Parent->use_begin(), E = Parent->use_end();
+ I != E; ++I) {
+
+ if (I.getUse().get() != Value)
+ continue;
+
+ if (I->getOpcode() == Opcode)
+ return *I;
+ }
+ return 0;
+}
+
+/// This transforms the control flow intrinsics to get the branch destination as
+/// last parameter, also switches branch target with BR if the need arise
+SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
+ SelectionDAG &DAG) const {
+
+ DebugLoc DL = BRCOND.getDebugLoc();
+
+ SDNode *Intr = BRCOND.getOperand(1).getNode();
+ SDValue Target = BRCOND.getOperand(2);
+ SDNode *BR = 0;
+
+ if (Intr->getOpcode() == ISD::SETCC) {
+ // As long as we negate the condition everything is fine
+ SDNode *SetCC = Intr;
+ assert(SetCC->getConstantOperandVal(1) == 1);
+ assert(cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get() ==
+ ISD::SETNE);
+ Intr = SetCC->getOperand(0).getNode();
+
+ } else {
+ // Get the target from BR if we don't negate the condition
+ BR = findUser(BRCOND, ISD::BR);
+ Target = BR->getOperand(1);
+ }
+
+ assert(Intr->getOpcode() == ISD::INTRINSIC_W_CHAIN);
+
+ // Build the result and
+ SmallVector<EVT, 4> Res;
+ for (unsigned i = 1, e = Intr->getNumValues(); i != e; ++i)
+ Res.push_back(Intr->getValueType(i));
+
+ // operands of the new intrinsic call
+ SmallVector<SDValue, 4> Ops;
+ Ops.push_back(BRCOND.getOperand(0));
+ for (unsigned i = 1, e = Intr->getNumOperands(); i != e; ++i)
+ Ops.push_back(Intr->getOperand(i));
+ Ops.push_back(Target);
+
+ // build the new intrinsic call
+ SDNode *Result = DAG.getNode(
+ Res.size() > 1 ? ISD::INTRINSIC_W_CHAIN : ISD::INTRINSIC_VOID, DL,
+ DAG.getVTList(Res.data(), Res.size()), Ops.data(), Ops.size()).getNode();
+
+ if (BR) {
+ // Give the branch instruction our target
+ SDValue Ops[] = {
+ BR->getOperand(0),
+ BRCOND.getOperand(2)
+ };
+ DAG.MorphNodeTo(BR, ISD::BR, BR->getVTList(), Ops, 2);
+ }
+
+ SDValue Chain = SDValue(Result, Result->getNumValues() - 1);
+
+ // Copy the intrinsic results to registers
+ for (unsigned i = 1, e = Intr->getNumValues() - 1; i != e; ++i) {
+ SDNode *CopyToReg = findUser(SDValue(Intr, i), ISD::CopyToReg);
+ if (!CopyToReg)
+ continue;
+
+ Chain = DAG.getCopyToReg(
+ Chain, DL,
+ CopyToReg->getOperand(1),
+ SDValue(Result, i - 1),
+ SDValue());
+
+ DAG.ReplaceAllUsesWith(SDValue(CopyToReg, 0), CopyToReg->getOperand(0));
+ }
+
+ // Remove the old intrinsic from the chain
+ DAG.ReplaceAllUsesOfValueWith(
+ SDValue(Intr, Intr->getNumValues() - 1),
+ Intr->getOperand(0));
+
+ return Chain;
}
SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
projects / oota-llvm.git / blobdiff