1 files changed, 170 insertions, 130 deletions

diff --git a/contrib/llvm/lib/Target/R600/R600MachineScheduler.cpp b/contrib/llvm/lib/Target/R600/R600MachineScheduler.cpp
index a777142a9e70..da2a4d862e7d 100644
--- a/contrib/llvm/lib/Target/R600/R600MachineScheduler.cpp
+++ b/contrib/llvm/lib/Target/R600/R600MachineScheduler.cpp
@@ -9,19 +9,17 @@
 //
 /// \file
 /// \brief R600 Machine Scheduler interface
-// TODO: Scheduling is optimised for VLIW4 arch, modify it to support TRANS slot
 //
 //===----------------------------------------------------------------------===//
 
 #define DEBUG_TYPE "misched"
 
 #include "R600MachineScheduler.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/Pass.h"
 #include "llvm/PassManager.h"
 #include "llvm/Support/raw_ostream.h"
-#include <set>
 
 using namespace llvm;
 
@@ -30,54 +28,80 @@ void R600SchedStrategy::initialize(ScheduleDAGMI *dag) {
   DAG = dag;
   TII = static_cast<const R600InstrInfo*>(DAG->TII);
   TRI = static_cast<const R600RegisterInfo*>(DAG->TRI);
+  VLIW5 = !DAG->MF.getTarget().getSubtarget<AMDGPUSubtarget>().hasCaymanISA();
   MRI = &DAG->MRI;
-  Available[IDAlu]->clear();
-  Available[IDFetch]->clear();
-  Available[IDOther]->clear();
   CurInstKind = IDOther;
   CurEmitted = 0;
-  OccupedSlotsMask = 15;
+  OccupedSlotsMask = 31;
   InstKindLimit[IDAlu] = TII->getMaxAlusPerClause();
-
+  InstKindLimit[IDOther] = 32;
 
   const AMDGPUSubtarget &ST = DAG->TM.getSubtarget<AMDGPUSubtarget>();
-  if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD5XXX) {
-    InstKindLimit[IDFetch] = 7; // 8 minus 1 for security
-  } else {
-    InstKindLimit[IDFetch] = 15; // 16 minus 1 for security
-  }
+  InstKindLimit[IDFetch] = ST.getTexVTXClauseSize();
+  AluInstCount = 0;
+  FetchInstCount = 0;
 }
 
-void R600SchedStrategy::MoveUnits(ReadyQueue *QSrc, ReadyQueue *QDst)
+void R600SchedStrategy::MoveUnits(std::vector<SUnit *> &QSrc,
+                                  std::vector<SUnit *> &QDst)
 {
-  if (QSrc->empty())
-    return;
-  for (ReadyQueue::iterator I = QSrc->begin(),
-      E = QSrc->end(); I != E; ++I) {
-    (*I)->NodeQueueId &= ~QSrc->getID();
-    QDst->push(*I);
-  }
-  QSrc->clear();
+  QDst.insert(QDst.end(), QSrc.begin(), QSrc.end());
+  QSrc.clear();
+}
+
+static
+unsigned getWFCountLimitedByGPR(unsigned GPRCount) {
+  assert (GPRCount && "GPRCount cannot be 0");
+  return 248 / GPRCount;
 }
 
 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]) ||
-      (Available[CurInstKind]->empty());
-  bool AllowSwitchFromAlu = (CurEmitted > InstKindLimit[CurInstKind]) &&
-      (!Available[IDFetch]->empty() || !Available[IDOther]->empty());
-
-  if ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
-      (!AllowSwitchFromAlu && CurInstKind == IDAlu)) {
+  bool AllowSwitchToAlu = (CurEmitted >= InstKindLimit[CurInstKind]) ||
+      (Available[CurInstKind].empty());
+  bool AllowSwitchFromAlu = (CurEmitted >= InstKindLimit[CurInstKind]) &&
+      (!Available[IDFetch].empty() || !Available[IDOther].empty());
+
+  if (CurInstKind == IDAlu && !Available[IDFetch].empty()) {
+    // We use the heuristic provided by AMD Accelerated Parallel Processing
+    // OpenCL Programming Guide :
+    // The approx. number of WF that allows TEX inst to hide ALU inst is :
+    // 500 (cycles for TEX) / (AluFetchRatio * 8 (cycles for ALU))
+    float ALUFetchRationEstimate = 
+        (AluInstCount + AvailablesAluCount() + Pending[IDAlu].size()) /
+        (FetchInstCount + Available[IDFetch].size());
+    unsigned NeededWF = 62.5f / ALUFetchRationEstimate;
+    DEBUG( dbgs() << NeededWF << " approx. Wavefronts Required\n" );
+    // We assume the local GPR requirements to be "dominated" by the requirement
+    // of the TEX clause (which consumes 128 bits regs) ; ALU inst before and
+    // after TEX are indeed likely to consume or generate values from/for the
+    // TEX clause.
+    // Available[IDFetch].size() * 2 : GPRs required in the Fetch clause
+    // We assume that fetch instructions are either TnXYZW = TEX TnXYZW (need
+    // one GPR) or TmXYZW = TnXYZW (need 2 GPR).
+    // (TODO : use RegisterPressure)
+    // If we are going too use too many GPR, we flush Fetch instruction to lower
+    // register pressure on 128 bits regs.
+    unsigned NearRegisterRequirement = 2 * Available[IDFetch].size();
+    if (NeededWF > getWFCountLimitedByGPR(NearRegisterRequirement))
+      AllowSwitchFromAlu = true;
+  }
+
+  if (!SU && ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
+      (!AllowSwitchFromAlu && CurInstKind == IDAlu))) {
     // try to pick ALU
     SU = pickAlu();
+    if (!SU && !PhysicalRegCopy.empty()) {
+      SU = PhysicalRegCopy.front();
+      PhysicalRegCopy.erase(PhysicalRegCopy.begin());
+    }
     if (SU) {
-      if (CurEmitted >  InstKindLimit[IDAlu])
+      if (CurEmitted >= InstKindLimit[IDAlu])
         CurEmitted = 0;
       NextInstKind = IDAlu;
     }
@@ -99,14 +123,10 @@ SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
 
   DEBUG(
       if (SU) {
-        dbgs() << "picked node: ";
+        dbgs() << " ** Pick node **\n";
         SU->dump(DAG);
       } else {
-        dbgs() << "NO NODE ";
-        for (int i = 0; i < IDLast; ++i) {
-          Available[i]->dump();
-          Pending[i]->dump();
-        }
+        dbgs() << "NO NODE \n";
         for (unsigned i = 0; i < DAG->SUnits.size(); i++) {
           const SUnit &S = DAG->SUnits[i];
           if (!S.isScheduled)
@@ -119,19 +139,16 @@ 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)
-      OccupedSlotsMask = 15;
+      OccupedSlotsMask |= 31;
     CurEmitted = 0;
     CurInstKind = NextInstKind;
   }
 
   if (CurInstKind == IDAlu) {
+    AluInstCount ++;
     switch (getAluKind(SU)) {
     case AluT_XYZW:
       CurEmitted += 4;
@@ -157,20 +174,37 @@ void R600SchedStrategy::schedNode(SUnit *SU, bool IsTopNode) {
 
   if (CurInstKind != IDFetch) {
     MoveUnits(Pending[IDFetch], Available[IDFetch]);
-  }
-  MoveUnits(Pending[IDOther], Available[IDOther]);
+  } else
+    FetchInstCount++;
 }
 
-void R600SchedStrategy::releaseTopNode(SUnit *SU) {
-  int IK = getInstKind(SU);
+static bool
+isPhysicalRegCopy(MachineInstr *MI) {
+  if (MI->getOpcode() != AMDGPU::COPY)
+    return false;
 
-  DEBUG(dbgs() << IK << " <= ");
-  DEBUG(SU->dump(DAG));
+  return !TargetRegisterInfo::isVirtualRegister(MI->getOperand(1).getReg());
+}
 
-  Pending[IK]->push(SU);
+void R600SchedStrategy::releaseTopNode(SUnit *SU) {
+  DEBUG(dbgs() << "Top Releasing ";SU->dump(DAG););
 }
 
 void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
+  DEBUG(dbgs() << "Bottom Releasing ";SU->dump(DAG););
+  if (isPhysicalRegCopy(SU->getInstr())) {
+    PhysicalRegCopy.push_back(SU);
+    return;
+  }
+
+  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,
@@ -185,18 +219,19 @@ bool R600SchedStrategy::regBelongsToClass(unsigned Reg,
 R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
   MachineInstr *MI = SU->getInstr();
 
+  if (TII->isTransOnly(MI))
+    return AluTrans;
+
     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;
       }
@@ -205,10 +240,18 @@ R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
     }
 
     // Does the instruction take a whole IG ?
+    // XXX: Is it possible to add a helper function in R600InstrInfo that can
+    // be used here and in R600PacketizerList::isSoloInstruction() ?
     if(TII->isVector(*MI) ||
         TII->isCubeOp(MI->getOpcode()) ||
-        TII->isReductionOp(MI->getOpcode()))
+        TII->isReductionOp(MI->getOpcode()) ||
+        MI->getOpcode() == AMDGPU::GROUP_BARRIER) {
       return AluT_XYZW;
+    }
+
+    if (TII->isLDSInstr(MI->getOpcode())) {
+      return AluT_X;
+    }
 
     // Is the result already assigned to a channel ?
     unsigned DestSubReg = MI->getOperand(0).getSubReg();
@@ -239,6 +282,10 @@ R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
     if (regBelongsToClass(DestReg, &AMDGPU::R600_Reg128RegClass))
       return AluT_XYZW;
 
+    // LDS src registers cannot be used in the Trans slot.
+    if (TII->readsLDSSrcReg(MI))
+      return AluT_XYZW;
+
     return AluAny;
 
 }
@@ -246,57 +293,39 @@ R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
 int R600SchedStrategy::getInstKind(SUnit* SU) {
   int Opcode = SU->getInstr()->getOpcode();
 
+  if (TII->usesTextureCache(Opcode) || TII->usesVertexCache(Opcode))
+    return IDFetch;
+
   if (TII->isALUInstr(Opcode)) {
     return IDAlu;
   }
 
   switch (Opcode) {
+  case AMDGPU::PRED_X:
   case AMDGPU::COPY:
   case AMDGPU::CONST_COPY:
   case AMDGPU::INTERP_PAIR_XY:
   case AMDGPU::INTERP_PAIR_ZW:
   case AMDGPU::INTERP_VEC_LOAD:
-  case AMDGPU::DOT4_eg_pseudo:
-  case AMDGPU::DOT4_r600_pseudo:
+  case AMDGPU::DOT_4:
     return IDAlu;
-  case AMDGPU::TEX_VTX_CONSTBUF:
-  case AMDGPU::TEX_VTX_TEXBUF:
-  case AMDGPU::TEX_LD:
-  case AMDGPU::TEX_GET_TEXTURE_RESINFO:
-  case AMDGPU::TEX_GET_GRADIENTS_H:
-  case AMDGPU::TEX_GET_GRADIENTS_V:
-  case AMDGPU::TEX_SET_GRADIENTS_H:
-  case AMDGPU::TEX_SET_GRADIENTS_V:
-  case AMDGPU::TEX_SAMPLE:
-  case AMDGPU::TEX_SAMPLE_C:
-  case AMDGPU::TEX_SAMPLE_L:
-  case AMDGPU::TEX_SAMPLE_C_L:
-  case AMDGPU::TEX_SAMPLE_LB:
-  case AMDGPU::TEX_SAMPLE_C_LB:
-  case AMDGPU::TEX_SAMPLE_G:
-  case AMDGPU::TEX_SAMPLE_C_G:
-  case AMDGPU::TXD:
-  case AMDGPU::TXD_SHADOW:
-    return IDFetch;
   default:
-    DEBUG(
-        dbgs() << "other inst: ";
-        SU->dump(DAG);
-    );
     return IDOther;
   }
 }
 
-SUnit *R600SchedStrategy::PopInst(std::multiset<SUnit *, CompareSUnit> &Q) {
+SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
   if (Q.empty())
     return NULL;
-  for (std::set<SUnit *, CompareSUnit>::iterator It = Q.begin(), E = Q.end();
+  for (std::vector<SUnit *>::reverse_iterator It = Q.rbegin(), E = Q.rend();
       It != E; ++It) {
     SUnit *SU = *It;
     InstructionsGroupCandidate.push_back(SU->getInstr());
-    if (TII->canBundle(InstructionsGroupCandidate)) {
+    if (TII->fitsConstReadLimitations(InstructionsGroupCandidate)
+        && (!AnyALU || !TII->isVectorOnly(SU->getInstr()))
+    ) {
       InstructionsGroupCandidate.pop_back();
-      Q.erase(It);
+      Q.erase((It + 1).base());
       return SU;
     } else {
       InstructionsGroupCandidate.pop_back();
@@ -306,33 +335,37 @@ SUnit *R600SchedStrategy::PopInst(std::multiset<SUnit *, CompareSUnit> &Q) {
 }
 
 void R600SchedStrategy::LoadAlu() {
-  ReadyQueue *QSrc = Pending[IDAlu];
-  for (ReadyQueue::iterator I = QSrc->begin(),
-        E = QSrc->end(); I != E; ++I) {
-      (*I)->NodeQueueId &= ~QSrc->getID();
-      AluKind AK = getAluKind(*I);
-      AvailableAlus[AK].insert(*I);
-    }
-    QSrc->clear();
+  std::vector<SUnit *> &QSrc = Pending[IDAlu];
+  for (unsigned i = 0, e = QSrc.size(); i < e; ++i) {
+    AluKind AK = getAluKind(QSrc[i]);
+    AvailableAlus[AK].push_back(QSrc[i]);
+  }
+  QSrc.clear();
 }
 
 void R600SchedStrategy::PrepareNextSlot() {
   DEBUG(dbgs() << "New Slot\n");
   assert (OccupedSlotsMask && "Slot wasn't filled");
   OccupedSlotsMask = 0;
+//  if (HwGen == AMDGPUSubtarget::NORTHERN_ISLANDS)
+//    OccupedSlotsMask |= 16;
   InstructionsGroupCandidate.clear();
   LoadAlu();
 }
 
 void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
-  unsigned DestReg = MI->getOperand(0).getReg();
+  int DstIndex = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst);
+  if (DstIndex == -1) {
+    return;
+  }
+  unsigned DestReg = MI->getOperand(DstIndex).getReg();
   // PressureRegister crashes if an operand is def and used in the same inst
   // and we try to constraint its regclass
   for (MachineInstr::mop_iterator It = MI->operands_begin(),
       E = MI->operands_end(); It != E; ++It) {
     MachineOperand &MO = *It;
     if (MO.isReg() && !MO.isDef() &&
-        MO.getReg() == MI->getOperand(0).getReg())
+        MO.getReg() == DestReg)
       return;
   }
   // Constrains the regclass of DestReg to assign it to Slot
@@ -352,53 +385,60 @@ void R600SchedStrategy::AssignSlot(MachineInstr* MI, unsigned Slot) {
   }
 }
 
-SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot) {
+SUnit *R600SchedStrategy::AttemptFillSlot(unsigned Slot, bool AnyAlu) {
   static const AluKind IndexToID[] = {AluT_X, AluT_Y, AluT_Z, AluT_W};
-  SUnit *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]]);
-  SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny]);
-  if (!UnslotedSU) {
+  SUnit *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]], AnyAlu);
+  if (SlotedSU)
     return SlotedSU;
-  } else if (!SlotedSU) {
+  SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny], AnyAlu);
+  if (UnslotedSU)
     AssignSlot(UnslotedSU->getInstr(), Slot);
-    return UnslotedSU;
-  } else {
-    //Determine which one to pick (the lesser one)
-    if (CompareSUnit()(SlotedSU, UnslotedSU)) {
-      AvailableAlus[AluAny].insert(UnslotedSU);
-      return SlotedSU;
-    } else {
-      AvailableAlus[IndexToID[Slot]].insert(SlotedSU);
-      AssignSlot(UnslotedSU->getInstr(), Slot);
-      return UnslotedSU;
-    }
-  }
+  return UnslotedSU;
 }
 
-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();
+unsigned R600SchedStrategy::AvailablesAluCount() const {
+  return AvailableAlus[AluAny].size() + AvailableAlus[AluT_XYZW].size() +
+      AvailableAlus[AluT_X].size() + AvailableAlus[AluT_Y].size() +
+      AvailableAlus[AluT_Z].size() + AvailableAlus[AluT_W].size() +
+      AvailableAlus[AluTrans].size() + AvailableAlus[AluDiscarded].size() +
+      AvailableAlus[AluPredX].size();
 }
 
 SUnit* R600SchedStrategy::pickAlu() {
-  while (!isAvailablesAluEmpty()) {
+  while (AvailablesAluCount() || !Pending[IDAlu].empty()) {
     if (!OccupedSlotsMask) {
+      // Bottom up scheduling : predX must comes first
+      if (!AvailableAlus[AluPredX].empty()) {
+        OccupedSlotsMask |= 31;
+        return PopInst(AvailableAlus[AluPredX], false);
+      }
       // Flush physical reg copies (RA will discard them)
       if (!AvailableAlus[AluDiscarded].empty()) {
-        OccupedSlotsMask = 15;
-        return PopInst(AvailableAlus[AluDiscarded]);
+        OccupedSlotsMask |= 31;
+        return PopInst(AvailableAlus[AluDiscarded], false);
       }
       // If there is a T_XYZW alu available, use it
       if (!AvailableAlus[AluT_XYZW].empty()) {
-        OccupedSlotsMask = 15;
-        return PopInst(AvailableAlus[AluT_XYZW]);
+        OccupedSlotsMask |= 15;
+        return PopInst(AvailableAlus[AluT_XYZW], false);
+      }
+    }
+    bool TransSlotOccuped = OccupedSlotsMask & 16;
+    if (!TransSlotOccuped && VLIW5) {
+      if (!AvailableAlus[AluTrans].empty()) {
+        OccupedSlotsMask |= 16;
+        return PopInst(AvailableAlus[AluTrans], false);
+      }
+      SUnit *SU = AttemptFillSlot(3, true);
+      if (SU) {
+        OccupedSlotsMask |= 16;
+        return SU;
       }
     }
-    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);
+        SUnit *SU = AttemptFillSlot(Chan, false);
         if (SU) {
           OccupedSlotsMask |= (1 << Chan);
           InstructionsGroupCandidate.push_back(SU->getInstr());
@@ -413,14 +453,14 @@ SUnit* R600SchedStrategy::pickAlu() {
 
 SUnit* R600SchedStrategy::pickOther(int QID) {
   SUnit *SU = 0;
-  ReadyQueue *AQ = Available[QID];
+  std::vector<SUnit *> &AQ = Available[QID];
 
-  if (AQ->empty()) {
+  if (AQ.empty()) {
     MoveUnits(Pending[QID], AQ);
   }
-  if (!AQ->empty()) {
-    SU = *AQ->begin();
-    AQ->remove(AQ->begin());
+  if (!AQ.empty()) {
+    SU = AQ.back();
+    AQ.resize(AQ.size() - 1);
   }
   return SU;
 }