#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#define GET_INSTRINFO_CTOR
+#define GET_INSTRINFO_CTOR_DTOR
#include "AMDGPUGenDFAPacketizer.inc"
using namespace llvm;
MachineBasicBlock::iterator MI, DebugLoc DL,
unsigned DestReg, unsigned SrcReg,
bool KillSrc) const {
- if (AMDGPU::R600_Reg128RegClass.contains(DestReg)
- && AMDGPU::R600_Reg128RegClass.contains(SrcReg)) {
- for (unsigned I = 0; I < 4; I++) {
+ unsigned VectorComponents = 0;
+ if (AMDGPU::R600_Reg128RegClass.contains(DestReg) &&
+ AMDGPU::R600_Reg128RegClass.contains(SrcReg)) {
+ VectorComponents = 4;
+ } else if(AMDGPU::R600_Reg64RegClass.contains(DestReg) &&
+ AMDGPU::R600_Reg64RegClass.contains(SrcReg)) {
+ VectorComponents = 2;
+ }
+
+ if (VectorComponents > 0) {
+ for (unsigned I = 0; I < VectorComponents; I++) {
unsigned SubRegIndex = RI.getSubRegFromChannel(I);
buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
RI.getSubReg(DestReg, SubRegIndex),
RegState::Define | RegState::Implicit);
}
} else {
-
- // We can't copy vec4 registers
- assert(!AMDGPU::R600_Reg128RegClass.contains(DestReg)
- && !AMDGPU::R600_Reg128RegClass.contains(SrcReg));
-
MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
DestReg, SrcReg);
NewMI->getOperand(getOperandIdx(*NewMI, AMDGPU::OpName::src0))
}
}
-MachineInstr * R600InstrInfo::getMovImmInstr(MachineFunction *MF,
- unsigned DstReg, int64_t Imm) const {
- MachineInstr * MI = MF->CreateMachineInstr(get(AMDGPU::MOV), DebugLoc());
- MachineInstrBuilder MIB(*MF, MI);
- MIB.addReg(DstReg, RegState::Define);
- MIB.addReg(AMDGPU::ALU_LITERAL_X);
- MIB.addImm(Imm);
- MIB.addReg(0); // PREDICATE_BIT
-
- return MI;
+/// \returns true if \p MBBI can be moved into a new basic.
+bool R600InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator MBBI) const {
+ for (MachineInstr::const_mop_iterator I = MBBI->operands_begin(),
+ E = MBBI->operands_end(); I != E; ++I) {
+ if (I->isReg() && !TargetRegisterInfo::isVirtualRegister(I->getReg()) &&
+ I->isUse() && RI.isPhysRegLiveAcrossClauses(I->getReg()))
+ return false;
+ }
+ return true;
}
unsigned R600InstrInfo::getIEQOpcode() const {
unsigned TargetFlags = get(Opcode).TSFlags;
return ((TargetFlags & R600_InstFlag::LDS_1A) |
- (TargetFlags & R600_InstFlag::LDS_1A1D));
+ (TargetFlags & R600_InstFlag::LDS_1A1D) |
+ (TargetFlags & R600_InstFlag::LDS_1A2D));
+}
+
+bool R600InstrInfo::isLDSNoRetInstr(unsigned Opcode) const {
+ return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) == -1;
+}
+
+bool R600InstrInfo::isLDSRetInstr(unsigned Opcode) const {
+ return isLDSInstr(Opcode) && getOperandIdx(Opcode, AMDGPU::OpName::dst) != -1;
+}
+
+bool R600InstrInfo::canBeConsideredALU(const MachineInstr *MI) const {
+ if (isALUInstr(MI->getOpcode()))
+ return true;
+ if (isVector(*MI) || isCubeOp(MI->getOpcode()))
+ return true;
+ switch (MI->getOpcode()) {
+ case AMDGPU::PRED_X:
+ case AMDGPU::INTERP_PAIR_XY:
+ case AMDGPU::INTERP_PAIR_ZW:
+ case AMDGPU::INTERP_VEC_LOAD:
+ case AMDGPU::COPY:
+ case AMDGPU::DOT_4:
+ return true;
+ default:
+ return false;
+ }
}
bool R600InstrInfo::isTransOnly(unsigned Opcode) const {
- return (get(Opcode).TSFlags & R600_InstFlag::TRANS_ONLY);
+ if (ST.hasCaymanISA())
+ return false;
+ return (get(Opcode).getSchedClass() == AMDGPU::Sched::TransALU);
}
bool R600InstrInfo::isTransOnly(const MachineInstr *MI) const {
return isTransOnly(MI->getOpcode());
}
+bool R600InstrInfo::isVectorOnly(unsigned Opcode) const {
+ return (get(Opcode).getSchedClass() == AMDGPU::Sched::VecALU);
+}
+
+bool R600InstrInfo::isVectorOnly(const MachineInstr *MI) const {
+ return isVectorOnly(MI->getOpcode());
+}
+
+bool R600InstrInfo::isExport(unsigned Opcode) const {
+ return (get(Opcode).TSFlags & R600_InstFlag::IS_EXPORT);
+}
+
bool R600InstrInfo::usesVertexCache(unsigned Opcode) const {
return ST.hasVertexCache() && IS_VTX(get(Opcode));
}
}
}
+bool R600InstrInfo::usesAddressRegister(MachineInstr *MI) const {
+ return MI->findRegisterUseOperandIdx(AMDGPU::AR_X) != -1;
+}
+
+bool R600InstrInfo::definesAddressRegister(MachineInstr *MI) const {
+ return MI->findRegisterDefOperandIdx(AMDGPU::AR_X) != -1;
+}
+
+bool R600InstrInfo::readsLDSSrcReg(const MachineInstr *MI) const {
+ if (!isALUInstr(MI->getOpcode())) {
+ return false;
+ }
+ for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
+ E = MI->operands_end(); I != E; ++I) {
+ if (!I->isReg() || !I->isUse() ||
+ TargetRegisterInfo::isVirtualRegister(I->getReg()))
+ continue;
+
+ if (AMDGPU::R600_LDS_SRC_REGRegClass.contains(I->getReg()))
+ return true;
+ }
+ return false;
+}
+
+int R600InstrInfo::getSrcIdx(unsigned Opcode, unsigned SrcNum) const {
+ static const unsigned OpTable[] = {
+ AMDGPU::OpName::src0,
+ AMDGPU::OpName::src1,
+ AMDGPU::OpName::src2
+ };
+
+ assert (SrcNum < 3);
+ return getOperandIdx(Opcode, OpTable[SrcNum]);
+}
+
+#define SRC_SEL_ROWS 11
+int R600InstrInfo::getSelIdx(unsigned Opcode, unsigned SrcIdx) const {
+ static const unsigned SrcSelTable[SRC_SEL_ROWS][2] = {
+ {AMDGPU::OpName::src0, AMDGPU::OpName::src0_sel},
+ {AMDGPU::OpName::src1, AMDGPU::OpName::src1_sel},
+ {AMDGPU::OpName::src2, AMDGPU::OpName::src2_sel},
+ {AMDGPU::OpName::src0_X, AMDGPU::OpName::src0_sel_X},
+ {AMDGPU::OpName::src0_Y, AMDGPU::OpName::src0_sel_Y},
+ {AMDGPU::OpName::src0_Z, AMDGPU::OpName::src0_sel_Z},
+ {AMDGPU::OpName::src0_W, AMDGPU::OpName::src0_sel_W},
+ {AMDGPU::OpName::src1_X, AMDGPU::OpName::src1_sel_X},
+ {AMDGPU::OpName::src1_Y, AMDGPU::OpName::src1_sel_Y},
+ {AMDGPU::OpName::src1_Z, AMDGPU::OpName::src1_sel_Z},
+ {AMDGPU::OpName::src1_W, AMDGPU::OpName::src1_sel_W}
+ };
+
+ for (unsigned i = 0; i < SRC_SEL_ROWS; ++i) {
+ if (getOperandIdx(Opcode, SrcSelTable[i][0]) == (int)SrcIdx) {
+ return getOperandIdx(Opcode, SrcSelTable[i][1]);
+ }
+ }
+ return -1;
+}
+#undef SRC_SEL_ROWS
+
SmallVector<std::pair<MachineOperand *, int64_t>, 3>
R600InstrInfo::getSrcs(MachineInstr *MI) const {
SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
static std::vector<std::pair<int, unsigned> >
Swizzle(std::vector<std::pair<int, unsigned> > Src,
R600InstrInfo::BankSwizzle Swz) {
+ if (Src[0] == Src[1])
+ Src[1].first = -1;
switch (Swz) {
case R600InstrInfo::ALU_VEC_012_SCL_210:
break;
isConstCompatible(R600InstrInfo::BankSwizzle TransSwz,
const std::vector<std::pair<int, unsigned> > &TransOps,
unsigned ConstCount) {
+ // TransALU can't read 3 constants
+ if (ConstCount > 2)
+ return false;
for (unsigned i = 0, e = TransOps.size(); i < e; ++i) {
const std::pair<int, unsigned> &Src = TransOps[i];
unsigned Cycle = getTransSwizzle(TransSwz, i);
R600InstrInfo::fitsConstReadLimitations(const std::vector<MachineInstr *> &MIs)
const {
std::vector<unsigned> Consts;
+ SmallSet<int64_t, 4> Literals;
for (unsigned i = 0, n = MIs.size(); i < n; i++) {
MachineInstr *MI = MIs[i];
if (!isALUInstr(MI->getOpcode()))
continue;
- const SmallVector<std::pair<MachineOperand *, int64_t>, 3> &Srcs =
+ const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Srcs =
getSrcs(MI);
for (unsigned j = 0, e = Srcs.size(); j < e; j++) {
std::pair<MachineOperand *, unsigned> Src = Srcs[j];
+ if (Src.first->getReg() == AMDGPU::ALU_LITERAL_X)
+ Literals.insert(Src.second);
+ if (Literals.size() > 4)
+ return false;
if (Src.first->getReg() == AMDGPU::ALU_CONST)
Consts.push_back(Src.second);
if (AMDGPU::R600_KC0RegClass.contains(Src.first->getReg()) ||
return Opcode == AMDGPU::JUMP || Opcode == AMDGPU::JUMP_COND;
}
+static bool isBranch(unsigned Opcode) {
+ return Opcode == AMDGPU::BRANCH || Opcode == AMDGPU::BRANCH_COND_i32 ||
+ Opcode == AMDGPU::BRANCH_COND_f32;
+}
+
bool
R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
MachineBasicBlock *&TBB,
return false;
--I;
}
+ // AMDGPU::BRANCH* instructions are only available after isel and are not
+ // handled
+ if (isBranch(I->getOpcode()))
+ return true;
if (!isJump(static_cast<MachineInstr *>(I)->getOpcode())) {
return false;
}
- // Get the last instruction in the block.
+ // Remove successive JUMP
+ while (I != MBB.begin() && std::prev(I)->getOpcode() == AMDGPU::JUMP) {
+ MachineBasicBlock::iterator PriorI = std::prev(I);
+ if (AllowModify)
+ I->removeFromParent();
+ I = PriorI;
+ }
MachineInstr *LastInst = I;
// If there is only one terminator instruction, process it.
It != E; ++It) {
if (It->getOpcode() == AMDGPU::CF_ALU ||
It->getOpcode() == AMDGPU::CF_ALU_PUSH_BEFORE)
- return llvm::prior(It.base());
+ return std::prev(It.base());
}
return MBB.end();
}
return true;
}
+ if (MI->getOpcode() == AMDGPU::DOT_4) {
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_X))
+ .setReg(Pred[2].getReg());
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Y))
+ .setReg(Pred[2].getReg());
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_Z))
+ .setReg(Pred[2].getReg());
+ MI->getOperand(getOperandIdx(*MI, AMDGPU::OpName::pred_sel_W))
+ .setReg(Pred[2].getReg());
+ MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI);
+ MIB.addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
+ return true;
+ }
+
if (PIdx != -1) {
MachineOperand &PMO = MI->getOperand(PIdx);
PMO.setReg(Pred[2].getReg());
return false;
}
+unsigned int R600InstrInfo::getPredicationCost(const MachineInstr *) const {
+ return 2;
+}
+
unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
const MachineInstr *MI,
unsigned *PredCost) const {
return 2;
}
-int R600InstrInfo::getIndirectIndexBegin(const MachineFunction &MF) const {
- const MachineRegisterInfo &MRI = MF.getRegInfo();
- const MachineFrameInfo *MFI = MF.getFrameInfo();
- int Offset = 0;
-
- if (MFI->getNumObjects() == 0) {
- return -1;
- }
-
- if (MRI.livein_empty()) {
- return 0;
- }
-
- for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(),
- LE = MRI.livein_end();
- LI != LE; ++LI) {
- Offset = std::max(Offset,
- GET_REG_INDEX(RI.getEncodingValue(LI->first)));
- }
-
- return Offset + 1;
-}
-
-int R600InstrInfo::getIndirectIndexEnd(const MachineFunction &MF) const {
- int Offset = 0;
- const MachineFrameInfo *MFI = MF.getFrameInfo();
-
- // Variable sized objects are not supported
- assert(!MFI->hasVarSizedObjects());
-
- if (MFI->getNumObjects() == 0) {
- return -1;
- }
-
- Offset = TM.getFrameLowering()->getFrameIndexOffset(MF, -1);
-
- return getIndirectIndexBegin(MF) + Offset;
-}
-
-std::vector<unsigned> R600InstrInfo::getIndirectReservedRegs(
+void R600InstrInfo::reserveIndirectRegisters(BitVector &Reserved,
const MachineFunction &MF) const {
const AMDGPUFrameLowering *TFL =
static_cast<const AMDGPUFrameLowering*>(TM.getFrameLowering());
- std::vector<unsigned> Regs;
unsigned StackWidth = TFL->getStackWidth(MF);
int End = getIndirectIndexEnd(MF);
- if (End == -1) {
- return Regs;
- }
+ if (End == -1)
+ return;
for (int Index = getIndirectIndexBegin(MF); Index <= End; ++Index) {
unsigned SuperReg = AMDGPU::R600_Reg128RegClass.getRegister(Index);
- Regs.push_back(SuperReg);
+ Reserved.set(SuperReg);
for (unsigned Chan = 0; Chan < StackWidth; ++Chan) {
unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister((4 * Index) + Chan);
- Regs.push_back(Reg);
+ Reserved.set(Reg);
}
}
- return Regs;
}
unsigned R600InstrInfo::calculateIndirectAddress(unsigned RegIndex,
return RegIndex;
}
-const TargetRegisterClass * R600InstrInfo::getIndirectAddrStoreRegClass(
- unsigned SourceReg) const {
- return &AMDGPU::R600_TReg32RegClass;
-}
-
-const TargetRegisterClass *R600InstrInfo::getIndirectAddrLoadRegClass() const {
- return &AMDGPU::TRegMemRegClass;
+const TargetRegisterClass *R600InstrInfo::getIndirectAddrRegClass() const {
+ return &AMDGPU::R600_TReg32_XRegClass;
}
MachineInstrBuilder R600InstrInfo::buildIndirectWrite(MachineBasicBlock *MBB,
return Mov;
}
-const TargetRegisterClass *R600InstrInfo::getSuperIndirectRegClass() const {
- return &AMDGPU::IndirectRegRegClass;
-}
-
unsigned R600InstrInfo::getMaxAlusPerClause() const {
return 115;
}
AMDGPU::OpName::src1_sel,
};
+ MachineOperand &MO = MI->getOperand(getOperandIdx(MI->getOpcode(),
+ getSlotedOps(AMDGPU::OpName::pred_sel, Slot)));
+ MIB->getOperand(getOperandIdx(Opcode, AMDGPU::OpName::pred_sel))
+ .setReg(MO.getReg());
+
for (unsigned i = 0; i < 14; i++) {
MachineOperand &MO = MI->getOperand(
getOperandIdx(MI->getOpcode(), getSlotedOps(Operands[i], Slot)));
return MovImm;
}
+MachineInstr *R600InstrInfo::buildMovInstr(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator I,
+ unsigned DstReg, unsigned SrcReg) const {
+ return buildDefaultInstruction(*MBB, I, AMDGPU::MOV, DstReg, SrcReg);
+}
+
int R600InstrInfo::getOperandIdx(const MachineInstr &MI, unsigned Op) const {
return getOperandIdx(MI.getOpcode(), Op);
}
projects / oota-llvm.git / blobdiff