//===----------------------------------------------------------------------===//
#include "R600ISelLowering.h"
+#include "AMDGPUFrameLowering.h"
+#include "AMDGPUIntrinsicInfo.h"
+#include "AMDGPUSubtarget.h"
#include "R600Defines.h"
#include "R600InstrInfo.h"
#include "R600MachineFunctionInfo.h"
+#include "llvm/Analysis/ValueTracking.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
setOperationAction(ISD::BR_CC, MVT::i32, Expand);
setOperationAction(ISD::BR_CC, MVT::f32, Expand);
+ setOperationAction(ISD::BRCOND, MVT::Other, Custom);
setOperationAction(ISD::FSUB, MVT::f32, Expand);
setOperationAction(ISD::SETCC, MVT::i32, Expand);
setOperationAction(ISD::SETCC, MVT::f32, Expand);
setOperationAction(ISD::FP_TO_UINT, MVT::i1, Custom);
+ setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
+ setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
setOperationAction(ISD::SELECT, MVT::i32, Expand);
setOperationAction(ISD::SELECT, MVT::f32, Expand);
setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
- setOperationAction(ISD::SELECT, MVT::v2f32, Expand);
setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
- setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
// Expand sign extension of vectors
if (!Subtarget->hasBFE())
setOperationAction(ISD::LOAD, MVT::v4i32, Custom);
setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i32, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f32, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
+
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i32, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f32, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
+
setTargetDAGCombine(ISD::FP_ROUND);
setTargetDAGCombine(ISD::FP_TO_SINT);
setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
setTargetDAGCombine(ISD::SELECT_CC);
setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
+ setOperationAction(ISD::SUB, MVT::i64, Expand);
+
+ // These should be replaced by UDVIREM, but it does not happen automatically
+ // during Type Legalization
+ setOperationAction(ISD::UDIV, MVT::i64, Custom);
+ setOperationAction(ISD::UREM, MVT::i64, Custom);
+ setOperationAction(ISD::SDIV, MVT::i64, Custom);
+ setOperationAction(ISD::SREM, MVT::i64, Custom);
+
+ // We don't have 64-bit shifts. Thus we need either SHX i64 or SHX_PARTS i32
+ // to be Legal/Custom in order to avoid library calls.
+ setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
+ setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
+ setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
+
setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
+ const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
+ for (MVT VT : ScalarIntVTs) {
+ setOperationAction(ISD::ADDC, VT, Expand);
+ setOperationAction(ISD::SUBC, VT, Expand);
+ setOperationAction(ISD::ADDE, VT, Expand);
+ setOperationAction(ISD::SUBE, VT, Expand);
+ }
+
setBooleanContents(ZeroOrNegativeOneBooleanContent);
setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
setSchedulingPreference(Sched::Source);
MachineRegisterInfo &MRI = MF->getRegInfo();
MachineBasicBlock::iterator I = *MI;
const R600InstrInfo *TII =
- static_cast<const R600InstrInfo*>(MF->getTarget().getInstrInfo());
+ static_cast<const R600InstrInfo *>(MF->getSubtarget().getInstrInfo());
switch (MI->getOpcode()) {
default:
int DstIdx = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst);
assert(DstIdx != -1);
MachineInstrBuilder NewMI;
- if (!MRI.use_empty(MI->getOperand(DstIdx).getReg()))
+ // FIXME: getLDSNoRetOp method only handles LDS_1A1D LDS ops. Add
+ // LDS_1A2D support and remove this special case.
+ if (!MRI.use_empty(MI->getOperand(DstIdx).getReg()) ||
+ MI->getOpcode() == AMDGPU::LDS_CMPST_RET)
return BB;
NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
switch (Op.getOpcode()) {
default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
+ case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
+ case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
+ case ISD::SHL_PARTS: return LowerSHLParts(Op, DAG);
+ case ISD::SRA_PARTS:
+ case ISD::SRL_PARTS: return LowerSRXParts(Op, DAG);
case ISD::FCOS:
case ISD::FSIN: return LowerTrig(Op, DAG);
case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
case ISD::STORE: return LowerSTORE(Op, DAG);
- case ISD::LOAD: return LowerLOAD(Op, DAG);
+ case ISD::LOAD: {
+ SDValue Result = LowerLOAD(Op, DAG);
+ assert((!Result.getNode() ||
+ Result.getNode()->getNumValues() == 2) &&
+ "Load should return a value and a chain");
+ return Result;
+ }
+
+ case ISD::BRCOND: return LowerBRCOND(Op, DAG);
case ISD::GlobalAddress: return LowerGlobalAddress(MFI, Op, DAG);
case ISD::INTRINSIC_VOID: {
SDValue Chain = Op.getOperand(0);
DAG.getConstant(2, MVT::i32), // SWZ_Z
DAG.getConstant(3, MVT::i32) // SWZ_W
};
- return DAG.getNode(AMDGPUISD::EXPORT, SDLoc(Op), Op.getValueType(),
- Args, 8);
+ return DAG.getNode(AMDGPUISD::EXPORT, SDLoc(Op), Op.getValueType(), Args);
}
// default for switch(IntrinsicID)
MachineSDNode *interp;
if (ijb < 0) {
const MachineFunction &MF = DAG.getMachineFunction();
- const R600InstrInfo *TII =
- static_cast<const R600InstrInfo*>(MF.getTarget().getInstrInfo());
+ const R600InstrInfo *TII = static_cast<const R600InstrInfo *>(
+ MF.getSubtarget().getInstrInfo());
interp = DAG.getMachineNode(AMDGPU::INTERP_VEC_LOAD, DL,
MVT::v4f32, DAG.getTargetConstant(slot / 4 , MVT::i32));
return DAG.getTargetExtractSubreg(
Op.getOperand(9),
Op.getOperand(10)
};
- return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs, 19);
+ return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, DL, MVT::v4f32, TexArgs);
}
case AMDGPUIntrinsic::AMDGPU_dp4: {
SDValue Args[8] = {
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, Op.getOperand(2),
DAG.getConstant(3, MVT::i32))
};
- return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args, 8);
+ return DAG.getNode(AMDGPUISD::DOT4, DL, MVT::f32, Args);
}
case Intrinsic::r600_read_ngroups_x:
case Intrinsic::r600_read_local_size_z:
return LowerImplicitParameter(DAG, VT, DL, 8);
+ case Intrinsic::AMDGPU_read_workdim:
+ return LowerImplicitParameter(DAG, VT, DL, MFI->ABIArgOffset / 4);
+
case Intrinsic::r600_read_tgid_x:
return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
AMDGPU::T1_X, VT);
case Intrinsic::r600_read_tidig_z:
return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
AMDGPU::T0_Z, VT);
+ case Intrinsic::AMDGPU_rsq:
+ // XXX - I'm assuming SI's RSQ_LEGACY matches R600's behavior.
+ return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
}
// break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
break;
default:
AMDGPUTargetLowering::ReplaceNodeResults(N, Results, DAG);
return;
- case ISD::FP_TO_UINT: Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
- return;
- case ISD::LOAD: {
- SDNode *Node = LowerLOAD(SDValue(N, 0), DAG).getNode();
- Results.push_back(SDValue(Node, 0));
- Results.push_back(SDValue(Node, 1));
- // XXX: LLVM seems not to replace Chain Value inside CustomWidenLowerNode
- // function
- DAG.ReplaceAllUsesOfValueWith(SDValue(N,1), SDValue(Node, 1));
+ case ISD::FP_TO_UINT:
+ if (N->getValueType(0) == MVT::i1) {
+ Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
+ return;
+ }
+ // Fall-through. Since we don't care about out of bounds values
+ // we can use FP_TO_SINT for uints too. The DAGLegalizer code for uint
+ // considers some extra cases which are not necessary here.
+ case ISD::FP_TO_SINT: {
+ SDValue Result;
+ if (expandFP_TO_SINT(N, Result, DAG))
+ Results.push_back(Result);
return;
}
- case ISD::STORE:
- SDNode *Node = LowerSTORE(SDValue(N, 0), DAG).getNode();
- Results.push_back(SDValue(Node, 0));
- return;
+ case ISD::UDIV: {
+ SDValue Op = SDValue(N, 0);
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
+ N->getOperand(0), N->getOperand(1));
+ Results.push_back(UDIVREM);
+ break;
+ }
+ case ISD::UREM: {
+ SDValue Op = SDValue(N, 0);
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
+ N->getOperand(0), N->getOperand(1));
+ Results.push_back(UDIVREM.getValue(1));
+ break;
+ }
+ case ISD::SDIV: {
+ SDValue Op = SDValue(N, 0);
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ SDValue SDIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(VT, VT),
+ N->getOperand(0), N->getOperand(1));
+ Results.push_back(SDIVREM);
+ break;
+ }
+ case ISD::SREM: {
+ SDValue Op = SDValue(N, 0);
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ SDValue SDIVREM = DAG.getNode(ISD::SDIVREM, DL, DAG.getVTList(VT, VT),
+ N->getOperand(0), N->getOperand(1));
+ Results.push_back(SDIVREM.getValue(1));
+ break;
+ }
+ case ISD::SDIVREM: {
+ SDValue Op = SDValue(N, 1);
+ SDValue RES = LowerSDIVREM(Op, DAG);
+ Results.push_back(RES);
+ Results.push_back(RES.getValue(1));
+ break;
+ }
+ case ISD::UDIVREM: {
+ SDValue Op = SDValue(N, 0);
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
+
+ SDValue one = DAG.getConstant(1, HalfVT);
+ SDValue zero = DAG.getConstant(0, HalfVT);
+
+ //HiLo split
+ SDValue LHS = N->getOperand(0);
+ SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
+ SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
+
+ SDValue RHS = N->getOperand(1);
+ SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
+ SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
+
+ // Get Speculative values
+ SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
+ SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
+
+ SDValue REM_Hi = zero;
+ SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
+
+ SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
+ SDValue DIV_Lo = zero;
+
+ const unsigned halfBitWidth = HalfVT.getSizeInBits();
+
+ for (unsigned i = 0; i < halfBitWidth; ++i) {
+ SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
+ // Get Value of high bit
+ SDValue HBit;
+ if (halfBitWidth == 32 && Subtarget->hasBFE()) {
+ HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
+ } else {
+ HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
+ HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
+ }
+
+ SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
+ DAG.getConstant(halfBitWidth - 1, HalfVT));
+ REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
+ REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
+
+ REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
+ REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
+
+
+ SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
+
+ SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
+ SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
+
+ DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
+
+ // Update REM
+
+ SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
+
+ REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
+ REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
+ REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
+ }
+
+ SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
+ SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
+ Results.push_back(DIV);
+ Results.push_back(REM);
+ break;
}
+ }
+}
+
+SDValue R600TargetLowering::vectorToVerticalVector(SelectionDAG &DAG,
+ SDValue Vector) const {
+
+ SDLoc DL(Vector);
+ EVT VecVT = Vector.getValueType();
+ EVT EltVT = VecVT.getVectorElementType();
+ SmallVector<SDValue, 8> Args;
+
+ for (unsigned i = 0, e = VecVT.getVectorNumElements();
+ i != e; ++i) {
+ Args.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT,
+ Vector, DAG.getConstant(i, getVectorIdxTy())));
+ }
+
+ return DAG.getNode(AMDGPUISD::BUILD_VERTICAL_VECTOR, DL, VecVT, Args);
+}
+
+SDValue R600TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
+ SelectionDAG &DAG) const {
+
+ SDLoc DL(Op);
+ SDValue Vector = Op.getOperand(0);
+ SDValue Index = Op.getOperand(1);
+
+ if (isa<ConstantSDNode>(Index) ||
+ Vector.getOpcode() == AMDGPUISD::BUILD_VERTICAL_VECTOR)
+ return Op;
+
+ Vector = vectorToVerticalVector(DAG, Vector);
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Op.getValueType(),
+ Vector, Index);
+}
+
+SDValue R600TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
+ SelectionDAG &DAG) const {
+ SDLoc DL(Op);
+ SDValue Vector = Op.getOperand(0);
+ SDValue Value = Op.getOperand(1);
+ SDValue Index = Op.getOperand(2);
+
+ if (isa<ConstantSDNode>(Index) ||
+ Vector.getOpcode() == AMDGPUISD::BUILD_VERTICAL_VECTOR)
+ return Op;
+
+ Vector = vectorToVerticalVector(DAG, Vector);
+ SDValue Insert = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, Op.getValueType(),
+ Vector, Value, Index);
+ return vectorToVerticalVector(DAG, Insert);
}
SDValue R600TargetLowering::LowerTrig(SDValue Op, SelectionDAG &DAG) const {
DAG.getConstantFP(3.14159265359, MVT::f32));
}
+SDValue R600TargetLowering::LowerSHLParts(SDValue Op, SelectionDAG &DAG) const {
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+
+ SDValue Lo = Op.getOperand(0);
+ SDValue Hi = Op.getOperand(1);
+ SDValue Shift = Op.getOperand(2);
+ SDValue Zero = DAG.getConstant(0, VT);
+ SDValue One = DAG.getConstant(1, VT);
+
+ SDValue Width = DAG.getConstant(VT.getSizeInBits(), VT);
+ SDValue Width1 = DAG.getConstant(VT.getSizeInBits() - 1, VT);
+ SDValue BigShift = DAG.getNode(ISD::SUB, DL, VT, Shift, Width);
+ SDValue CompShift = DAG.getNode(ISD::SUB, DL, VT, Width1, Shift);
+
+ // The dance around Width1 is necessary for 0 special case.
+ // Without it the CompShift might be 32, producing incorrect results in
+ // Overflow. So we do the shift in two steps, the alternative is to
+ // add a conditional to filter the special case.
+
+ SDValue Overflow = DAG.getNode(ISD::SRL, DL, VT, Lo, CompShift);
+ Overflow = DAG.getNode(ISD::SRL, DL, VT, Overflow, One);
+
+ SDValue HiSmall = DAG.getNode(ISD::SHL, DL, VT, Hi, Shift);
+ HiSmall = DAG.getNode(ISD::OR, DL, VT, HiSmall, Overflow);
+ SDValue LoSmall = DAG.getNode(ISD::SHL, DL, VT, Lo, Shift);
+
+ SDValue HiBig = DAG.getNode(ISD::SHL, DL, VT, Lo, BigShift);
+ SDValue LoBig = Zero;
+
+ Hi = DAG.getSelectCC(DL, Shift, Width, HiSmall, HiBig, ISD::SETULT);
+ Lo = DAG.getSelectCC(DL, Shift, Width, LoSmall, LoBig, ISD::SETULT);
+
+ return DAG.getNode(ISD::MERGE_VALUES, DL, DAG.getVTList(VT,VT), Lo, Hi);
+}
+
+SDValue R600TargetLowering::LowerSRXParts(SDValue Op, SelectionDAG &DAG) const {
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+
+ SDValue Lo = Op.getOperand(0);
+ SDValue Hi = Op.getOperand(1);
+ SDValue Shift = Op.getOperand(2);
+ SDValue Zero = DAG.getConstant(0, VT);
+ SDValue One = DAG.getConstant(1, VT);
+
+ const bool SRA = Op.getOpcode() == ISD::SRA_PARTS;
+
+ SDValue Width = DAG.getConstant(VT.getSizeInBits(), VT);
+ SDValue Width1 = DAG.getConstant(VT.getSizeInBits() - 1, VT);
+ SDValue BigShift = DAG.getNode(ISD::SUB, DL, VT, Shift, Width);
+ SDValue CompShift = DAG.getNode(ISD::SUB, DL, VT, Width1, Shift);
+
+ // The dance around Width1 is necessary for 0 special case.
+ // Without it the CompShift might be 32, producing incorrect results in
+ // Overflow. So we do the shift in two steps, the alternative is to
+ // add a conditional to filter the special case.
+
+ SDValue Overflow = DAG.getNode(ISD::SHL, DL, VT, Hi, CompShift);
+ Overflow = DAG.getNode(ISD::SHL, DL, VT, Overflow, One);
+
+ SDValue HiSmall = DAG.getNode(SRA ? ISD::SRA : ISD::SRL, DL, VT, Hi, Shift);
+ SDValue LoSmall = DAG.getNode(ISD::SRL, DL, VT, Lo, Shift);
+ LoSmall = DAG.getNode(ISD::OR, DL, VT, LoSmall, Overflow);
+
+ SDValue LoBig = DAG.getNode(SRA ? ISD::SRA : ISD::SRL, DL, VT, Hi, BigShift);
+ SDValue HiBig = SRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, Width1) : Zero;
+
+ Hi = DAG.getSelectCC(DL, Shift, Width, HiSmall, HiBig, ISD::SETULT);
+ Lo = DAG.getSelectCC(DL, Shift, Width, LoSmall, LoBig, ISD::SETULT);
+
+ return DAG.getNode(ISD::MERGE_VALUES, DL, DAG.getVTList(VT,VT), Lo, Hi);
+}
+
SDValue R600TargetLowering::LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const {
return DAG.getNode(
ISD::SETCC,
return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
}
-
- // Possible Min/Max pattern
- SDValue MinMax = LowerMinMax(Op, DAG);
- if (MinMax.getNode()) {
- return MinMax;
- }
-
// If we make it this for it means we have no native instructions to handle
// this SELECT_CC, so we must lower it.
SDValue HWTrue, HWFalse;
DAG.getConstant(0, MVT::i32),
Mask
};
- SDValue Input = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, Src, 4);
+ SDValue Input = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v4i32, Src);
SDValue Args[3] = { Chain, Input, DWordAddr };
return DAG.getMemIntrinsicNode(AMDGPUISD::STORE_MSKOR, DL,
- Op->getVTList(), Args, 3, MemVT,
+ Op->getVTList(), Args, MemVT,
StoreNode->getMemOperand());
} else if (Ptr->getOpcode() != AMDGPUISD::DWORDADDR &&
Value.getValueType().bitsGE(MVT::i32)) {
// Lowering for indirect addressing
const MachineFunction &MF = DAG.getMachineFunction();
- const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
- getTargetMachine().getFrameLowering());
+ const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
+ getTargetMachine().getSubtargetImpl()->getFrameLowering());
unsigned StackWidth = TFL->getStackWidth(MF);
Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
if (ValueVT.isVector()) {
unsigned NumElemVT = ValueVT.getVectorNumElements();
EVT ElemVT = ValueVT.getVectorElementType();
- SDValue Stores[4];
+ SmallVector<SDValue, 4> Stores(NumElemVT);
assert(NumElemVT >= StackWidth && "Stack width cannot be greater than "
"vector width in load");
Chain, Elem, Ptr,
DAG.getTargetConstant(Channel, MVT::i32));
}
- Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores, NumElemVT);
+ Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Stores);
} else {
if (ValueVT == MVT::i8) {
Value = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Value);
Ret,
Chain
};
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
}
+ // Lower loads constant address space global variable loads
+ if (LoadNode->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
+ isa<GlobalVariable>(
+ GetUnderlyingObject(LoadNode->getMemOperand()->getValue()))) {
+
+ SDValue Ptr = DAG.getZExtOrTrunc(LoadNode->getBasePtr(), DL,
+ getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
+ Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
+ DAG.getConstant(2, MVT::i32));
+ return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op->getVTList(),
+ LoadNode->getChain(), Ptr,
+ DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
+ }
if (LoadNode->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS && VT.isVector()) {
SDValue MergedValues[2] = {
- SplitVectorLoad(Op, DAG),
+ ScalarizeVectorLoad(Op, DAG),
Chain
};
- return DAG.getMergeValues(MergedValues, 2, DL);
+ return DAG.getMergeValues(MergedValues, DL);
}
int ConstantBlock = ConstantAddressBlock(LoadNode->getAddressSpace());
NewVT = VT;
NumElements = VT.getVectorNumElements();
}
- Result = DAG.getNode(ISD::BUILD_VECTOR, DL, NewVT, Slots, NumElements);
+ Result = DAG.getNode(ISD::BUILD_VECTOR, DL, NewVT,
+ makeArrayRef(Slots, NumElements));
} else {
// non-constant ptr can't be folded, keeps it as a v4f32 load
Result = DAG.getNode(AMDGPUISD::CONST_ADDRESS, DL, MVT::v4i32,
Result,
Chain
};
- return DAG.getMergeValues(MergedValues, 2, DL);
+ return DAG.getMergeValues(MergedValues, DL);
}
// For most operations returning SDValue() will result in the node being
LoadNode->getPointerInfo(), MemVT,
LoadNode->isVolatile(),
LoadNode->isNonTemporal(),
+ LoadNode->isInvariant(),
LoadNode->getAlignment());
SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, NewLoad, ShiftAmount);
SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, Shl, ShiftAmount);
SDValue MergedValues[2] = { Sra, Chain };
- return DAG.getMergeValues(MergedValues, 2, DL);
+ return DAG.getMergeValues(MergedValues, DL);
}
if (LoadNode->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
// Lowering for indirect addressing
const MachineFunction &MF = DAG.getMachineFunction();
- const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering*>(
- getTargetMachine().getFrameLowering());
+ const AMDGPUFrameLowering *TFL = static_cast<const AMDGPUFrameLowering *>(
+ getTargetMachine().getSubtargetImpl()->getFrameLowering());
unsigned StackWidth = TFL->getStackWidth(MF);
Ptr = stackPtrToRegIndex(Ptr, StackWidth, DAG);
Loads[i] = DAG.getUNDEF(ElemVT);
}
EVT TargetVT = EVT::getVectorVT(*DAG.getContext(), ElemVT, 4);
- LoweredLoad = DAG.getNode(ISD::BUILD_VECTOR, DL, TargetVT, Loads, 4);
+ LoweredLoad = DAG.getNode(ISD::BUILD_VECTOR, DL, TargetVT, Loads);
} else {
LoweredLoad = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, VT,
Chain, Ptr,
Chain
};
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
+}
+
+SDValue R600TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
+ SDValue Chain = Op.getOperand(0);
+ SDValue Cond = Op.getOperand(1);
+ SDValue Jump = Op.getOperand(2);
+
+ return DAG.getNode(AMDGPUISD::BRANCH_COND, SDLoc(Op), Op.getValueType(),
+ Chain, Jump, Cond);
}
/// XXX Only kernel functions are supported, so we can assume for now that
SDLoc DL, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const {
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
- getTargetMachine(), ArgLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
+ *DAG.getContext());
MachineFunction &MF = DAG.getMachineFunction();
- unsigned ShaderType = MF.getInfo<R600MachineFunctionInfo>()->ShaderType;
+ R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
SmallVector<ISD::InputArg, 8> LocalIns;
for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
CCValAssign &VA = ArgLocs[i];
- EVT VT = Ins[i].VT;
- EVT MemVT = LocalIns[i].VT;
+ const ISD::InputArg &In = Ins[i];
+ EVT VT = In.VT;
+ EVT MemVT = VA.getLocVT();
+ if (!VT.isVector() && MemVT.isVector()) {
+ // Get load source type if scalarized.
+ MemVT = MemVT.getVectorElementType();
+ }
- if (ShaderType != ShaderType::COMPUTE) {
+ if (MFI->getShaderType() != ShaderType::COMPUTE) {
unsigned Reg = MF.addLiveIn(VA.getLocReg(), &AMDGPU::R600_Reg128RegClass);
SDValue Register = DAG.getCopyFromReg(Chain, DL, Reg, VT);
InVals.push_back(Register);
}
PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
- AMDGPUAS::CONSTANT_BUFFER_0);
+ AMDGPUAS::CONSTANT_BUFFER_0);
// i64 isn't a legal type, so the register type used ends up as i32, which
// isn't expected here. It attempts to create this sextload, but it ends up
// The first 36 bytes of the input buffer contains information about
// thread group and global sizes.
+ ISD::LoadExtType Ext = ISD::NON_EXTLOAD;
+ if (MemVT.getScalarSizeInBits() != VT.getScalarSizeInBits()) {
+ // FIXME: This should really check the extload type, but the handling of
+ // extload vector parameters seems to be broken.
- // FIXME: This should really check the extload type, but the handling of
- // extload vecto parameters seems to be broken.
- //ISD::LoadExtType Ext = Ins[i].Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
- ISD::LoadExtType Ext = ISD::SEXTLOAD;
- SDValue Arg = DAG.getExtLoad(Ext, DL, VT, Chain,
- DAG.getConstant(36 + VA.getLocMemOffset(), MVT::i32),
- MachinePointerInfo(UndefValue::get(PtrTy)),
- MemVT, false, false, 4);
+ // Ext = In.Flags.isSExt() ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
+ Ext = ISD::SEXTLOAD;
+ }
+
+ // Compute the offset from the value.
+ // XXX - I think PartOffset should give you this, but it seems to give the
+ // size of the register which isn't useful.
+
+ unsigned ValBase = ArgLocs[In.OrigArgIndex].getLocMemOffset();
+ unsigned PartOffset = VA.getLocMemOffset();
+ unsigned Offset = 36 + VA.getLocMemOffset();
+
+ MachinePointerInfo PtrInfo(UndefValue::get(PtrTy), PartOffset - ValBase);
+ SDValue Arg = DAG.getLoad(ISD::UNINDEXED, Ext, VT, DL, Chain,
+ DAG.getConstant(Offset, MVT::i32),
+ DAG.getUNDEF(MVT::i32),
+ PtrInfo,
+ MemVT, false, true, true, 4);
// 4 is the preferred alignment for the CONSTANT memory space.
InVals.push_back(Arg);
+ MFI->ABIArgOffset = Offset + MemVT.getStoreSize();
}
return Chain;
}
}
return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
- VectorEntry.getValueType(), NewBldVec, 4);
+ VectorEntry.getValueType(), NewBldVec);
}
static SDValue ReorganizeVector(SelectionDAG &DAG, SDValue VectorEntry,
}
return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(VectorEntry),
- VectorEntry.getValueType(), NewBldVec, 4);
+ VectorEntry.getValueType(), NewBldVec);
}
}
// Return the new vector
- return DAG.getNode(ISD::BUILD_VECTOR, dl,
- VT, Ops.data(), Ops.size());
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
}
// Extract_vec (Build_vector) generated by custom lowering
}
case ISD::SELECT_CC: {
+ // Try common optimizations
+ SDValue Ret = AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
+ if (Ret.getNode())
+ return Ret;
+
// fold selectcc (selectcc x, y, a, b, cc), b, a, b, seteq ->
// selectcc x, y, a, b, inv(cc)
//
};
SDLoc DL(N);
NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[4], DAG);
- return DAG.getNode(AMDGPUISD::EXPORT, DL, N->getVTList(), NewArgs, 8);
+ return DAG.getNode(AMDGPUISD::EXPORT, DL, N->getVTList(), NewArgs);
}
case AMDGPUISD::TEXTURE_FETCH: {
SDValue Arg = N->getOperand(1);
};
NewArgs[1] = OptimizeSwizzle(N->getOperand(1), &NewArgs[2], DAG);
return DAG.getNode(AMDGPUISD::TEXTURE_FETCH, SDLoc(N), N->getVTList(),
- NewArgs, 19);
+ NewArgs);
}
}
- return SDValue();
+
+ return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
}
static bool
FoldOperand(SDNode *ParentNode, unsigned SrcIdx, SDValue &Src, SDValue &Neg,
SDValue &Abs, SDValue &Sel, SDValue &Imm, SelectionDAG &DAG) {
const R600InstrInfo *TII =
- static_cast<const R600InstrInfo *>(DAG.getTarget().getInstrInfo());
+ static_cast<const R600InstrInfo *>(DAG.getSubtarget().getInstrInfo());
if (!Src.isMachineOpcode())
return false;
switch (Src.getMachineOpcode()) {
TII->getOperandIdx(Opcode, AMDGPU::OpName::src1_W)
};
std::vector<unsigned> Consts;
- for (unsigned i = 0; i < sizeof(SrcIndices) / sizeof(int); i++) {
- int OtherSrcIdx = SrcIndices[i];
+ for (int OtherSrcIdx : SrcIndices) {
int OtherSelIdx = TII->getSelIdx(Opcode, OtherSrcIdx);
if (OtherSrcIdx < 0 || OtherSelIdx < 0)
continue;
if (RegisterSDNode *Reg =
dyn_cast<RegisterSDNode>(ParentNode->getOperand(OtherSrcIdx))) {
if (Reg->getReg() == AMDGPU::ALU_CONST) {
- ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(
- ParentNode->getOperand(OtherSelIdx));
+ ConstantSDNode *Cst
+ = cast<ConstantSDNode>(ParentNode->getOperand(OtherSelIdx));
Consts.push_back(Cst->getZExtValue());
}
}
}
- ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(CstOffset);
+ ConstantSDNode *Cst = cast<ConstantSDNode>(CstOffset);
Consts.push_back(Cst->getZExtValue());
if (!TII->fitsConstReadLimitations(Consts)) {
return false;
SDNode *R600TargetLowering::PostISelFolding(MachineSDNode *Node,
SelectionDAG &DAG) const {
const R600InstrInfo *TII =
- static_cast<const R600InstrInfo *>(DAG.getTarget().getInstrInfo());
+ static_cast<const R600InstrInfo *>(DAG.getSubtarget().getInstrInfo());
if (!Node->isMachineOpcode())
return Node;
unsigned Opcode = Node->getMachineOpcode();
SDValue FakeOp;
std::vector<SDValue> Ops;
- for(SDNode::op_iterator I = Node->op_begin(), E = Node->op_end();
- I != E; ++I)
- Ops.push_back(*I);
+ for (const SDUse &I : Node->ops())
+ Ops.push_back(I);
if (Opcode == AMDGPU::DOT_4) {
int OperandIdx[] = {
projects / oota-llvm.git / blobdiff