return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
case Intrinsic::AMDGPU_rsq_clamped:
- return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
+ if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
+ Type *Type = VT.getTypeForEVT(*DAG.getContext());
+ APFloat Max = APFloat::getLargest(Type->getFltSemantics());
+ APFloat Min = APFloat::getLargest(Type->getFltSemantics(), true);
+
+ SDValue Rsq = DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
+ SDValue Tmp = DAG.getNode(ISD::FMINNUM, DL, VT, Rsq,
+ DAG.getConstantFP(Max, VT));
+ return DAG.getNode(ISD::FMAXNUM, DL, VT, Tmp,
+ DAG.getConstantFP(Min, VT));
+ } else {
+ return DAG.getNode(AMDGPUISD::RSQ_CLAMPED, DL, VT, Op.getOperand(1));
+ }
case Intrinsic::AMDGPU_ldexp:
return DAG.getNode(AMDGPUISD::LDEXP, DL, VT, Op.getOperand(1),
case ISD::SETOLT:
case ISD::SETLE:
case ISD::SETLT: {
+ if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
+ break;
+
// We need to permute the operands to get the correct NaN behavior. The
// selected operand is the second one based on the failing compare with NaN,
// so permute it based on the compare type the hardware uses.
case ISD::SETOGE:
case ISD::SETUGT:
case ISD::SETOGT: {
+ if (Subtarget->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
+ break;
+
if (LHS == True)
return DAG.getNode(AMDGPUISD::FMAX_LEGACY, DL, VT, RHS, LHS);
return DAG.getNode(AMDGPUISD::FMIN_LEGACY, DL, VT, LHS, RHS);
return Dst;
}
+// Change the order of operands from (0, 1, 2) to (0, 2, 1)
+void SIInstrInfo::swapOperands(MachineBasicBlock::iterator Inst) const {
+ assert(Inst->getNumExplicitOperands() == 3);
+ MachineOperand Op1 = Inst->getOperand(1);
+ Inst->RemoveOperand(1);
+ Inst->addOperand(Op1);
+}
+
bool SIInstrInfo::isOperandLegal(const MachineInstr *MI, unsigned OpIdx,
const MachineOperand *MO) const {
const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
continue;
}
+ case AMDGPU::S_LSHL_B32:
+ if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
+ NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
+ swapOperands(Inst);
+ }
+ break;
+ case AMDGPU::S_ASHR_I32:
+ if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
+ NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
+ swapOperands(Inst);
+ }
+ break;
+ case AMDGPU::S_LSHR_B32:
+ if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
+ NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
+ swapOperands(Inst);
+ }
+ break;
+
case AMDGPU::S_BFE_U64:
case AMDGPU::S_BFM_B64:
llvm_unreachable("Moving this op to VALU not implemented");