From 76fc2d077f955174c14e658bf179620ef49dd792 Mon Sep 17 00:00:00 2001 From: Vincent Lejeune Date: Fri, 17 May 2013 16:50:56 +0000 Subject: R600: Use bottom up scheduling algorithm git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@182129 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/R600/R600MachineScheduler.cpp | 51 ++++++++++++++++++-------------- lib/Target/R600/R600MachineScheduler.h | 2 +- lib/Target/R600/R600RegisterInfo.cpp | 6 +++- lib/Target/R600/R600RegisterInfo.h | 3 ++ 4 files changed, 37 insertions(+), 25 deletions(-) (limited to 'lib') diff --git a/lib/Target/R600/R600MachineScheduler.cpp b/lib/Target/R600/R600MachineScheduler.cpp index aeb2674f4e..4c89124f46 100644 --- a/lib/Target/R600/R600MachineScheduler.cpp +++ b/lib/Target/R600/R600MachineScheduler.cpp @@ -34,7 +34,7 @@ void R600SchedStrategy::initialize(ScheduleDAGMI *dag) { CurEmitted = 0; OccupedSlotsMask = 15; InstKindLimit[IDAlu] = TII->getMaxAlusPerClause(); - + InstKindLimit[IDOther] = 32; const AMDGPUSubtarget &ST = DAG->TM.getSubtarget(); InstKindLimit[IDFetch] = ST.getTexVTXClauseSize(); @@ -49,12 +49,12 @@ void R600SchedStrategy::MoveUnits(std::vector &QSrc, SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) { SUnit *SU = 0; - IsTopNode = true; NextInstKind = IDOther; + IsTopNode = false; + // check if we might want to switch current clause type - bool AllowSwitchToAlu = (CurInstKind == IDOther) || - (CurEmitted >= InstKindLimit[CurInstKind]) || + bool AllowSwitchToAlu = (CurEmitted >= InstKindLimit[CurInstKind]) || (Available[CurInstKind].empty()); bool AllowSwitchFromAlu = (CurEmitted >= InstKindLimit[CurInstKind]) && (!Available[IDFetch].empty() || !Available[IDOther].empty()); @@ -86,10 +86,10 @@ SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) { DEBUG( if (SU) { - dbgs() << "picked node: "; + dbgs() << " ** Pick node **\n"; SU->dump(DAG); } else { - dbgs() << "NO NODE "; + dbgs() << "NO NODE \n"; for (unsigned i = 0; i < DAG->SUnits.size(); i++) { const SUnit &S = DAG->SUnits[i]; if (!S.isScheduled) @@ -103,9 +103,6 @@ SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) { void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) { - DEBUG(dbgs() << "scheduled: "); - DEBUG(SU->dump(DAG)); - if (NextInstKind != CurInstKind) { DEBUG(dbgs() << "Instruction Type Switch\n"); if (NextInstKind != IDAlu) @@ -141,19 +138,23 @@ void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) { if (CurInstKind != IDFetch) { MoveUnits(Pending[IDFetch], Available[IDFetch]); } - MoveUnits(Pending[IDOther], Available[IDOther]); } void R600SchedStrategy::releaseTopNode(SUnit *SU) { - int IK = getInstKind(SU); - - DEBUG(dbgs() << IK << " <= "); - DEBUG(SU->dump(DAG)); + DEBUG(dbgs() << "Top Releasing ";SU->dump(DAG);); - Pending[IK].push_back(SU); } void R600SchedStrategy::releaseBottomNode(SUnit *SU) { + DEBUG(dbgs() << "Bottom Releasing ";SU->dump(DAG);); + + int IK = getInstKind(SU); + // There is no export clause, we can schedule one as soon as its ready + if (IK == IDOther) + Available[IDOther].push_back(SU); + else + Pending[IK].push_back(SU); + } bool R600SchedStrategy::regBelongsToClass(unsigned Reg, @@ -169,18 +170,15 @@ R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const { MachineInstr *MI = SU->getInstr(); switch (MI->getOpcode()) { + case AMDGPU::PRED_X: + return AluPredX; case AMDGPU::INTERP_PAIR_XY: case AMDGPU::INTERP_PAIR_ZW: case AMDGPU::INTERP_VEC_LOAD: case AMDGPU::DOT_4: return AluT_XYZW; case AMDGPU::COPY: - if (TargetRegisterInfo::isPhysicalRegister(MI->getOperand(1).getReg())) { - // %vregX = COPY Tn_X is likely to be discarded in favor of an - // assignement of Tn_X to %vregX, don't considers it in scheduling - return AluDiscarded; - } - else if (MI->getOperand(1).isUndef()) { + if (MI->getOperand(1).isUndef()) { // MI will become a KILL, don't considers it in scheduling return AluDiscarded; } @@ -238,6 +236,7 @@ int R600SchedStrategy::getInstKind(SUnit* SU) { } switch (Opcode) { + case AMDGPU::PRED_X: case AMDGPU::COPY: case AMDGPU::CONST_COPY: case AMDGPU::INTERP_PAIR_XY: @@ -328,12 +327,18 @@ bool R600SchedStrategy::isAvailablesAluEmpty() const { return Pending[IDAlu].empty() && AvailableAlus[AluAny].empty() && AvailableAlus[AluT_XYZW].empty() && AvailableAlus[AluT_X].empty() && AvailableAlus[AluT_Y].empty() && AvailableAlus[AluT_Z].empty() && - AvailableAlus[AluT_W].empty() && AvailableAlus[AluDiscarded].empty(); + AvailableAlus[AluT_W].empty() && AvailableAlus[AluDiscarded].empty() && + AvailableAlus[AluPredX].empty(); } SUnit* R600SchedStrategy::pickAlu() { while (!isAvailablesAluEmpty()) { if (!OccupedSlotsMask) { + // Bottom up scheduling : predX must comes first + if (!AvailableAlus[AluPredX].empty()) { + OccupedSlotsMask = 15; + return PopInst(AvailableAlus[AluPredX]); + } // Flush physical reg copies (RA will discard them) if (!AvailableAlus[AluDiscarded].empty()) { OccupedSlotsMask = 15; @@ -345,7 +350,7 @@ SUnit* R600SchedStrategy::pickAlu() { return PopInst(AvailableAlus[AluT_XYZW]); } } - for (unsigned Chan = 0; Chan < 4; ++Chan) { + for (int Chan = 3; Chan > -1; --Chan) { bool isOccupied = OccupedSlotsMask & (1 << Chan); if (!isOccupied) { SUnit *SU = AttemptFillSlot(Chan); diff --git a/lib/Target/R600/R600MachineScheduler.h b/lib/Target/R600/R600MachineScheduler.h index c82ee49c78..1db877db85 100644 --- a/lib/Target/R600/R600MachineScheduler.h +++ b/lib/Target/R600/R600MachineScheduler.h @@ -45,13 +45,13 @@ class R600SchedStrategy : public MachineSchedStrategy { AluT_Z, AluT_W, AluT_XYZW, + AluPredX, AluDiscarded, // LLVM Instructions that are going to be eliminated AluLast }; std::vector Available[IDLast], Pending[IDLast]; std::vector AvailableAlus[AluLast]; - std::vector FakeCopy; InstKind CurInstKind; int CurEmitted; diff --git a/lib/Target/R600/R600RegisterInfo.cpp b/lib/Target/R600/R600RegisterInfo.cpp index bbd7995d7d..7d13420a31 100644 --- a/lib/Target/R600/R600RegisterInfo.cpp +++ b/lib/Target/R600/R600RegisterInfo.cpp @@ -25,7 +25,7 @@ R600RegisterInfo::R600RegisterInfo(AMDGPUTargetMachine &tm, : AMDGPURegisterInfo(tm, tii), TM(tm), TII(tii) - { } + { RCW.RegWeight = 0; RCW.WeightLimit = 0;} BitVector R600RegisterInfo::getReservedRegs(const MachineFunction &MF) const { BitVector Reserved(getNumRegs()); @@ -97,3 +97,7 @@ unsigned R600RegisterInfo::getSubRegFromChannel(unsigned Channel) const { } } +const RegClassWeight &R600RegisterInfo::getRegClassWeight( + const TargetRegisterClass *RC) const { + return RCW; +} diff --git a/lib/Target/R600/R600RegisterInfo.h b/lib/Target/R600/R600RegisterInfo.h index f9ca918f24..1270a1eea6 100644 --- a/lib/Target/R600/R600RegisterInfo.h +++ b/lib/Target/R600/R600RegisterInfo.h @@ -26,6 +26,7 @@ class TargetInstrInfo; struct R600RegisterInfo : public AMDGPURegisterInfo { AMDGPUTargetMachine &TM; const TargetInstrInfo &TII; + RegClassWeight RCW; R600RegisterInfo(AMDGPUTargetMachine &tm, const TargetInstrInfo &tii); @@ -48,6 +49,8 @@ struct R600RegisterInfo : public AMDGPURegisterInfo { /// (e.g. getSubRegFromChannel(0) -> AMDGPU::sel_x) unsigned getSubRegFromChannel(unsigned Channel) const; + virtual const RegClassWeight &getRegClassWeight(const TargetRegisterClass *RC) const; + }; } // End namespace llvm -- cgit v1.2.3