R600: Non vector only instruction can be scheduled on trans unit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@187514 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 60 insertions, 19 deletions

diff --git a/lib/Target/R600/R600InstrInfo.cpp b/lib/Target/R600/R600InstrInfo.cpp
index 83f8e71fbf..7ef3d85123 100644
--- a/lib/Target/R600/R600InstrInfo.cpp
+++ b/lib/Target/R600/R600InstrInfo.cpp
@@ -469,6 +469,9 @@ static bool
 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);
diff --git a/lib/Target/R600/R600MachineScheduler.cpp b/lib/Target/R600/R600MachineScheduler.cpp
index 0dc0365926..0499dd52d9 100644
--- a/lib/Target/R600/R600MachineScheduler.cpp
+++ b/lib/Target/R600/R600MachineScheduler.cpp
@@ -9,7 +9,6 @@
 //
 /// \file
 /// \brief R600 Machine Scheduler interface
-// TODO: Scheduling is optimised for VLIW4 arch, modify it to support TRANS slot
 //
 //===----------------------------------------------------------------------===//
 
@@ -29,6 +28,7 @@ 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;
   CurInstKind = IDOther;
   CurEmitted = 0;
@@ -342,14 +342,16 @@ int R600SchedStrategy::getInstKind(SUnit* SU) {
   }
 }
 
-SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q) {
+SUnit *R600SchedStrategy::PopInst(std::vector<SUnit *> &Q, bool AnyALU) {
   if (Q.empty())
     return NULL;
   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->fitsConstReadLimitations(InstructionsGroupCandidate)) {
+    if (TII->fitsConstReadLimitations(InstructionsGroupCandidate)
+        && (!AnyALU || !TII->isVectorOnly(SU->getInstr()))
+    ) {
       InstructionsGroupCandidate.pop_back();
       Q.erase((It + 1).base());
       return SU;
@@ -373,6 +375,8 @@ 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();
 }
@@ -409,12 +413,12 @@ 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 *SlotedSU = PopInst(AvailableAlus[IndexToID[Slot]], AnyAlu);
   if (SlotedSU)
     return SlotedSU;
-  SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny]);
+  SUnit *UnslotedSU = PopInst(AvailableAlus[AluAny], AnyAlu);
   if (UnslotedSU)
     AssignSlot(UnslotedSU->getInstr(), Slot);
   return UnslotedSU;
@@ -434,30 +438,35 @@ SUnit* R600SchedStrategy::pickAlu() {
       // Bottom up scheduling : predX must comes first
       if (!AvailableAlus[AluPredX].empty()) {
         OccupedSlotsMask |= 31;
-        return PopInst(AvailableAlus[AluPredX]);
+        return PopInst(AvailableAlus[AluPredX], false);
       }
       // Flush physical reg copies (RA will discard them)
       if (!AvailableAlus[AluDiscarded].empty()) {
         OccupedSlotsMask |= 31;
-        return PopInst(AvailableAlus[AluDiscarded]);
+        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]);
+        return PopInst(AvailableAlus[AluT_XYZW], false);
       }
     }
     bool TransSlotOccuped = OccupedSlotsMask & 16;
-    if (!TransSlotOccuped) {
+    if (!TransSlotOccuped && VLIW5) {
       if (!AvailableAlus[AluTrans].empty()) {
         OccupedSlotsMask |= 16;
-        return PopInst(AvailableAlus[AluTrans]);
+        return PopInst(AvailableAlus[AluTrans], false);
+      }
+      SUnit *SU = AttemptFillSlot(3, true);
+      if (SU) {
+        OccupedSlotsMask |= 16;
+        return SU;
       }
     }
     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());
diff --git a/lib/Target/R600/R600MachineScheduler.h b/lib/Target/R600/R600MachineScheduler.h
index f8965d8998..0a6f1204a4 100644
--- a/lib/Target/R600/R600MachineScheduler.h
+++ b/lib/Target/R600/R600MachineScheduler.h
@@ -84,15 +84,16 @@ public:
 
 private:
   std::vector<MachineInstr *> InstructionsGroupCandidate;
+  bool VLIW5;
 
   int getInstKind(SUnit *SU);
   bool regBelongsToClass(unsigned Reg, const TargetRegisterClass *RC) const;
   AluKind getAluKind(SUnit *SU) const;
   void LoadAlu();
   unsigned AvailablesAluCount() const;
-  SUnit *AttemptFillSlot (unsigned Slot);
+  SUnit *AttemptFillSlot (unsigned Slot, bool AnyAlu);
   void PrepareNextSlot();
-  SUnit *PopInst(std::vector<SUnit*> &Q);
+  SUnit *PopInst(std::vector<SUnit*> &Q, bool AnyALU);
 
   void AssignSlot(MachineInstr *MI, unsigned Slot);
   SUnit* pickAlu();
diff --git a/lib/Target/R600/R600Packetizer.cpp b/lib/Target/R600/R600Packetizer.cpp
index 5cf1fd3b66..6c70052b20 100644
--- a/lib/Target/R600/R600Packetizer.cpp
+++ b/lib/Target/R600/R600Packetizer.cpp
@@ -58,6 +58,8 @@ class R600PacketizerList : public VLIWPacketizerList {
 private:
   const R600InstrInfo *TII;
   const R600RegisterInfo &TRI;
+  bool VLIW5;
+  bool ConsideredInstUsesAlreadyWrittenVectorElement;
 
   unsigned getSlot(const MachineInstr *MI) const {
     return TRI.getHWRegChan(MI->getOperand(0).getReg());
@@ -74,7 +76,13 @@ private:
     MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
     if (I->isBundle())
       BI++;
+    int LastDstChan = -1;
     do {
+      bool isTrans = false;
+      int BISlot = getSlot(BI);
+      if (LastDstChan >= BISlot)
+        isTrans = true;
+      LastDstChan = BISlot;
       if (TII->isPredicated(BI))
         continue;
       int OperandIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::write);
@@ -85,7 +93,7 @@ private:
         continue;
       }
       unsigned Dst = BI->getOperand(DstIdx).getReg();
-      if (TII->isTransOnly(BI)) {
+      if (isTrans || TII->isTransOnly(BI)) {
         Result[Dst] = AMDGPU::PS;
         continue;
       }
@@ -142,10 +150,14 @@ public:
                         MachineDominatorTree &MDT)
   : VLIWPacketizerList(MF, MLI, MDT, true),
     TII (static_cast<const R600InstrInfo *>(MF.getTarget().getInstrInfo())),
-    TRI(TII->getRegisterInfo()) { }
+    TRI(TII->getRegisterInfo()) {
+    VLIW5 = !MF.getTarget().getSubtarget<AMDGPUSubtarget>().hasCaymanISA();
+  }
 
   // initPacketizerState - initialize some internal flags.
-  void initPacketizerState() { }
+  void initPacketizerState() {
+    ConsideredInstUsesAlreadyWrittenVectorElement = false;
+  }
 
   // ignorePseudoInstruction - Ignore bundling of pseudo instructions.
   bool ignorePseudoInstruction(MachineInstr *MI, MachineBasicBlock *MBB) {
@@ -172,8 +184,8 @@ public:
   // together.
   bool isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) {
     MachineInstr *MII = SUI->getInstr(), *MIJ = SUJ->getInstr();
-    if (getSlot(MII) <= getSlot(MIJ) && !TII->isTransOnly(MII))
-      return false;
+    if (getSlot(MII) == getSlot(MIJ))
+      ConsideredInstUsesAlreadyWrittenVectorElement = true;
     // Does MII and MIJ share the same pred_sel ?
     int OpI = TII->getOperandIdx(MII->getOpcode(), AMDGPU::OpName::pred_sel),
         OpJ = TII->getOperandIdx(MIJ->getOpcode(), AMDGPU::OpName::pred_sel);
@@ -211,6 +223,20 @@ public:
                                  std::vector<R600InstrInfo::BankSwizzle> &BS,
                                  bool &isTransSlot) {
     isTransSlot = TII->isTransOnly(MI);
+    assert (!isTransSlot || VLIW5);
+
+    // Is the dst reg sequence legal ?
+    if (!isTransSlot && !CurrentPacketMIs.empty()) {
+      if (getSlot(MI) <= getSlot(CurrentPacketMIs.back())) {
+        if (ConsideredInstUsesAlreadyWrittenVectorElement  &&
+            !TII->isVectorOnly(MI) && VLIW5) {
+          isTransSlot = true;
+          DEBUG(dbgs() << "Considering as Trans Inst :"; MI->dump(););
+        }
+        else
+          return false;
+      }
+    }
 
     // Are the Constants limitations met ?
     CurrentPacketMIs.push_back(MI);
@@ -278,6 +304,8 @@ public:
       return It;
     }
     endPacket(MI->getParent(), MI);
+    if (TII->isTransOnly(MI))
+      return MI;
     return VLIWPacketizerList::addToPacket(MI);
   }
 };