R600: Fix scheduling of instructions that use the LDS output queue

The LDS output queue is accessed via the OQAP register.  The OQAP
register cannot be live across clauses, so if value is written to the
output queue, it must be retrieved before the end of the clause.
With the machine scheduler, we cannot statisfy this constraint, because
it lacks proper alias analysis and it will mark some LDS accesses as
having a chain dependency on vertex fetches.  Since vertex fetches
require a new clauses, the dependency may end up spiltting OQAP uses and
defs so the end up in different clauses.  See the lds-output-queue.ll
test for a more detailed explanation.

To work around this issue, we now combine the LDS read and the OQAP
copy into one instruction and expand it after register allocation.

This patch also adds some checks to the EmitClauseMarker pass, so that
it doesn't end a clause with a value still in the output queue and
removes AR.X and OQAP handling from the scheduler (AR.X uses and defs
were already being expanded post-RA, so the scheduler will never see
them).

Reviewed-by: Vincent Lejeune <vljn at ovi.com>
llvm-svn: 194755

lib/Target/R600/R600EmitClauseMarkers.cpp

@@ -47,6 +47,11 @@ private:
       break;
     }
 
+    // These will be expanded to two ALU instructions in the
+    // ExpandSpecialInstructions pass.
+    if (TII->isLDSRetInstr(MI->getOpcode()))
+      return 2;
+
     if(TII->isVector(*MI) ||
         TII->isCubeOp(MI->getOpcode()) ||
         TII->isReductionOp(MI->getOpcode()))
@@ -106,8 +111,13 @@ private:
   }
 
   bool SubstituteKCacheBank(MachineInstr *MI,
-      std::vector<std::pair<unsigned, unsigned> > &CachedConsts) const {
+      std::vector<std::pair<unsigned, unsigned> > &CachedConsts,
+      bool UpdateInstr = true) const {
     std::vector<std::pair<unsigned, unsigned> > UsedKCache;
+
+    if (!TII->isALUInstr(MI->getOpcode()) && MI->getOpcode() != AMDGPU::DOT_4)
+      return true;
+
     const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Consts =
         TII->getSrcs(MI);
     assert((TII->isALUInstr(MI->getOpcode()) ||
@@ -140,6 +150,9 @@ private:
       return false;
     }
 
+    if (!UpdateInstr)
+      return true;
+
     for (unsigned i = 0, j = 0, n = Consts.size(); i < n; ++i) {
       if (Consts[i].first->getReg() != AMDGPU::ALU_CONST)
         continue;
@@ -160,6 +173,52 @@ private:
     return true;
   }
 
+  bool canClauseLocalKillFitInClause(
+                        unsigned AluInstCount,
+                        std::vector<std::pair<unsigned, unsigned> > KCacheBanks,
+                        MachineBasicBlock::iterator Def,
+                        MachineBasicBlock::iterator BBEnd) {
+    const R600RegisterInfo &TRI = TII->getRegisterInfo();
+    for (MachineInstr::const_mop_iterator
+           MOI = Def->operands_begin(),
+           MOE = Def->operands_end(); MOI != MOE; ++MOI) {
+      if (!MOI->isReg() || !MOI->isDef() ||
+          TRI.isPhysRegLiveAcrossClauses(MOI->getReg()))
+        continue;
+
+      // Def defines a clause local register, so check that its use will fit
+      // in the clause.
+      unsigned LastUseCount = 0;
+      for (MachineBasicBlock::iterator UseI = Def; UseI != BBEnd; ++UseI) {
+        AluInstCount += OccupiedDwords(UseI);
+        // Make sure we won't need to end the clause due to KCache limitations.
+        if (!SubstituteKCacheBank(UseI, KCacheBanks, false))
+          return false;
+
+        // We have reached the maximum instruction limit before finding the
+        // use that kills this register, so we cannot use this def in the
+        // current clause.
+        if (AluInstCount >= TII->getMaxAlusPerClause())
+          return false;
+
+        // Register kill flags have been cleared by the time we get to this
+        // pass, but it is safe to assume that all uses of this register
+        // occur in the same basic block as its definition, because
+        // it is illegal for the scheduler to schedule them in
+        // different blocks.
+        if (UseI->findRegisterUseOperandIdx(MOI->getReg()))
+          LastUseCount = AluInstCount;
+
+        if (UseI != Def && UseI->findRegisterDefOperandIdx(MOI->getReg()) != -1)
+          break;
+      }
+      if (LastUseCount)
+        return LastUseCount <= TII->getMaxAlusPerClause();
+      llvm_unreachable("Clause local register live at end of clause.");
+    }
+    return true;
+  }
+
   MachineBasicBlock::iterator
   MakeALUClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator I) {
     MachineBasicBlock::iterator ClauseHead = I;
@@ -198,11 +257,13 @@ private:
         I++;
         break;
       }
-      if (TII->isALUInstr(I->getOpcode()) &&
-          !SubstituteKCacheBank(I, KCacheBanks))
+
+      // If this instruction defines a clause local register, make sure
+      // its use can fit in this clause.
+      if (!canClauseLocalKillFitInClause(AluInstCount, KCacheBanks, I, E))
         break;
-      if (I->getOpcode() == AMDGPU::DOT_4 &&
-          !SubstituteKCacheBank(I, KCacheBanks))
+
+      if (!SubstituteKCacheBank(I, KCacheBanks))
         break;
       AluInstCount += OccupiedDwords(I);
     }

lib/Target/R600/R600ExpandSpecialInstrs.cpp

@@ -68,6 +68,23 @@ bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
       MachineInstr &MI = *I;
       I = llvm::next(I);
 
+      // Expand LDS_*_RET instructions
+      if (TII->isLDSRetInstr(MI.getOpcode())) {
+        int DstIdx = TII->getOperandIdx(MI.getOpcode(), AMDGPU::OpName::dst);
+        assert(DstIdx != -1);
+        MachineOperand &DstOp = MI.getOperand(DstIdx);
+        MachineInstr *Mov = TII->buildMovInstr(&MBB, I,
+                                               DstOp.getReg(), AMDGPU::OQAP);
+        DstOp.setReg(AMDGPU::OQAP);
+        int LDSPredSelIdx = TII->getOperandIdx(MI.getOpcode(),
+                                           AMDGPU::OpName::pred_sel);
+        int MovPredSelIdx = TII->getOperandIdx(Mov->getOpcode(),
+                                           AMDGPU::OpName::pred_sel);
+        // Copy the pred_sel bit
+        Mov->getOperand(MovPredSelIdx).setReg(
+            MI.getOperand(LDSPredSelIdx).getReg());
+      }
+
       switch (MI.getOpcode()) {
       default: break;
       // Expand PRED_X to one of the PRED_SET instructions.

lib/Target/R600/R600ISelLowering.cpp

@@ -134,21 +134,17 @@ MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
 
   switch (MI->getOpcode()) {
   default:
-    if (TII->isLDSInstr(MI->getOpcode()) &&
-        TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst) != -1) {
+    // Replace LDS_*_RET instruction that don't have any uses with the
+    // equivalent LDS_*_NORET instruction.
+    if (TII->isLDSRetInstr(MI->getOpcode())) {
       int DstIdx = TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::dst);
       assert(DstIdx != -1);
       MachineInstrBuilder NewMI;
-      if (!MRI.use_empty(MI->getOperand(DstIdx).getReg())) {
-        NewMI = BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()),
-                        AMDGPU::OQAP);
-        TII->buildDefaultInstruction(*BB, I, AMDGPU::MOV,
-                                     MI->getOperand(0).getReg(),
-                                     AMDGPU::OQAP);
-      } else {
-        NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
-                        TII->get(AMDGPU::getLDSNoRetOp(MI->getOpcode())));
-      }
+      if (!MRI.use_empty(MI->getOperand(DstIdx).getReg()))
+        return BB;
+
+      NewMI = BuildMI(*BB, I, BB->findDebugLoc(I),
+                      TII->get(AMDGPU::getLDSNoRetOp(MI->getOpcode())));
       for (unsigned i = 1, e = MI->getNumOperands(); i < e; ++i) {
         NewMI.addOperand(MI->getOperand(i));
       }

lib/Target/R600/R600InstrInfo.cpp

@@ -141,6 +141,14 @@ bool R600InstrInfo::isLDSInstr(unsigned Opcode) const {
           (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;

lib/Target/R600/R600InstrInfo.h

@@ -65,6 +65,8 @@ namespace llvm {
   bool isALUInstr(unsigned Opcode) const;
   bool hasInstrModifiers(unsigned Opcode) const;
   bool isLDSInstr(unsigned Opcode) const;
+  bool isLDSNoRetInstr(unsigned Opcode) const;
+  bool isLDSRetInstr(unsigned Opcode) const;
 
   /// \returns true if this \p Opcode represents an ALU instruction or an
   /// instruction that will be lowered in ExpandSpecialInstrs Pass.

lib/Target/R600/R600MachineScheduler.cpp

@@ -92,15 +92,6 @@ SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
       AllowSwitchFromAlu = true;
   }
 
-
-  // We want to scheduled AR defs as soon as possible to make sure they aren't
-  // put in a different ALU clause from their uses.
-  if (!SU && !UnscheduledARDefs.empty()) {
-      SU = UnscheduledARDefs[0];
-      UnscheduledARDefs.erase(UnscheduledARDefs.begin());
-      NextInstKind = IDAlu;
-  }
-
   if (!SU && ((AllowSwitchToAlu && CurInstKind != IDAlu) ||
       (!AllowSwitchFromAlu && CurInstKind == IDAlu))) {
     // try to pick ALU
@@ -130,15 +121,6 @@ SUnit* R600SchedStrategy::pickNode(bool &IsTopNode) {
       NextInstKind = IDOther;
   }
 
-  // We want to schedule the AR uses as late as possible to make sure that
-  // the AR defs have been released.
-  if (!SU && !UnscheduledARUses.empty()) {
-      SU = UnscheduledARUses[0];
-      UnscheduledARUses.erase(UnscheduledARUses.begin());
-      NextInstKind = IDAlu;
-  }
-
-
   DEBUG(
       if (SU) {
         dbgs() << " ** Pick node **\n";
@@ -217,20 +199,6 @@ void R600SchedStrategy::releaseBottomNode(SUnit *SU) {
 
   int IK = getInstKind(SU);
 
-  // Check for AR register defines
-  for (MachineInstr::const_mop_iterator I = SU->getInstr()->operands_begin(),
-                                        E = SU->getInstr()->operands_end();
-                                        I != E; ++I) {
-    if (I->isReg() && I->getReg() == AMDGPU::AR_X) {
-      if (I->isDef()) {
-        UnscheduledARDefs.push_back(SU);
-      } else {
-        UnscheduledARUses.push_back(SU);
-      }
-      return;
-    }
-  }
-
   // There is no export clause, we can schedule one as soon as its ready
   if (IK == IDOther)
     Available[IDOther].push_back(SU);

lib/Target/R600/R600MachineScheduler.h

@@ -53,8 +53,6 @@ class R600SchedStrategy : public MachineSchedStrategy {
 
   std::vector<SUnit *> Available[IDLast], Pending[IDLast];
   std::vector<SUnit *> AvailableAlus[AluLast];
-  std::vector<SUnit *> UnscheduledARDefs;
-  std::vector<SUnit *> UnscheduledARUses;
   std::vector<SUnit *> PhysicalRegCopy;
 
   InstKind CurInstKind;

lib/Target/R600/R600RegisterInfo.cpp

@@ -85,3 +85,16 @@ const RegClassWeight &R600RegisterInfo::getRegClassWeight(
   const TargetRegisterClass *RC) const {
   return RCW;
 }
+
+bool R600RegisterInfo::isPhysRegLiveAcrossClauses(unsigned Reg) const {
+  assert(!TargetRegisterInfo::isVirtualRegister(Reg));
+
+  switch (Reg) {
+  case AMDGPU::OQAP:
+  case AMDGPU::OQBP:
+  case AMDGPU::AR_X:
+    return false;
+  default:
+    return true;
+  }
+}

lib/Target/R600/R600RegisterInfo.h

@@ -47,6 +47,8 @@ struct R600RegisterInfo : public AMDGPURegisterInfo {
 
   virtual const RegClassWeight &getRegClassWeight(const TargetRegisterClass *RC) const;
 
+  // \returns true if \p Reg can be defined in one ALU caluse and used in another.
+  virtual bool isPhysRegLiveAcrossClauses(unsigned Reg) const;
 };
 
 } // End namespace llvm

test/CodeGen/R600/lds-output-queue.ll

@@ -0,0 +1,99 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood -verify-machineinstrs | FileCheck %s
+;
+; This test checks that the lds input queue will is empty at the end of
+; the ALU clause.
+
+; CHECK-LABEL: @lds_input_queue
+; CHECK: LDS_READ_RET * OQAP
+; CHECK-NOT: ALU clause
+; CHECK: MOV * T{{[0-9]\.[XYZW]}}, OQAP
+
+@local_mem = internal addrspace(3) unnamed_addr global [2 x i32] [i32 1, i32 2], align 4
+
+define void @lds_input_queue(i32 addrspace(1)* %out, i32 addrspace(1)* %in, i32 %index) {
+entry:
+  %0 = getelementptr inbounds [2 x i32] addrspace(3)* @local_mem, i32 0, i32 %index
+  %1 = load i32 addrspace(3)* %0
+  call void @llvm.AMDGPU.barrier.local()
+
+  ; This will start a new clause for the vertex fetch
+  %2 = load i32 addrspace(1)* %in
+  %3 = add i32 %1, %2
+  store i32 %3, i32 addrspace(1)* %out
+  ret void
+}
+
+declare void @llvm.AMDGPU.barrier.local()
+
+; The machine scheduler does not do proper alias analysis and assumes that
+; loads from global values (Note that a global value is different that a
+; value from global memory.  A global value is a value that is declared
+; outside of a function, it can reside in any address space) alias with
+; all other loads.
+;
+; This is a problem for scheduling the reads from the local data share (lds).
+; These reads are implemented using two instructions.  The first copies the
+; data from lds into the lds output queue, and the second moves the data from
+; the input queue into main memory.  These two instructions don't have to be
+; scheduled one after the other, but they do need to be scheduled in the same
+; clause.  The aliasing problem mentioned above causes problems when there is a
+; load from global memory which immediately follows a load from a global value that
+; has been declared in the local memory space:
+;
+;  %0 = getelementptr inbounds [2 x i32] addrspace(3)* @local_mem, i32 0, i32 %index
+;  %1 = load i32 addrspace(3)* %0
+;  %2 = load i32 addrspace(1)* %in
+;
+; The instruction selection phase will generate ISA that looks like this:
+; %OQAP = LDS_READ_RET
+; %vreg0 = MOV %OQAP
+; %vreg1 = VTX_READ_32
+; %vreg2 = ADD_INT %vreg1, %vreg0
+;
+; The bottom scheduler will schedule the two ALU instructions first:
+;
+; UNSCHEDULED:
+; %OQAP = LDS_READ_RET
+; %vreg1 = VTX_READ_32
+;
+; SCHEDULED:
+;
+; vreg0 = MOV %OQAP
+; vreg2 = ADD_INT %vreg1, %vreg2
+;
+; The lack of proper aliasing results in the local memory read (LDS_READ_RET)
+; to consider the global memory read (VTX_READ_32) has a chain dependency, so
+; the global memory read will always be scheduled first.  This will give us a
+; final program which looks like this:
+;
+; Alu clause:
+; %OQAP = LDS_READ_RET
+; VTX clause:
+; %vreg1 = VTX_READ_32
+; Alu clause:
+; vreg0 = MOV %OQAP
+; vreg2 = ADD_INT %vreg1, %vreg2
+;
+; This is an illegal program because the OQAP def and use know occur in
+; different ALU clauses.
+;
+; This test checks this scenario and makes sure it doesn't result in an
+; illegal program.  For now, we have fixed this issue by merging the
+; LDS_READ_RET and MOV together during instruction selection and then
+; expanding them after scheduling.  Once the scheduler has better alias
+; analysis, we should be able to keep these instructions sparate before
+; scheduling.
+;
+; CHECK-LABEL: @local_global_alias
+; CHECK: LDS_READ_RET
+; CHECK-NOT: ALU clause
+; CHECK MOV * T{{[0-9]\.[XYZW]}}, OQAP
+define void @local_global_alias(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+entry:
+  %0 = getelementptr inbounds [2 x i32] addrspace(3)* @local_mem, i32 0, i32 0
+  %1 = load i32 addrspace(3)* %0
+  %2 = load i32 addrspace(1)* %in
+  %3 = add i32 %2, %1
+  store i32 %3, i32 addrspace(1)* %out
+  ret void
+}

test/CodeGen/R600/local-memory-two-objects.ll

@@ -12,9 +12,11 @@
 ; SI-CHECK: .long 47180
 ; SI-CHECK-NEXT: .long 32768
 
-; Make sure the lds writes are using different addresses.
-; EG-CHECK: LDS_WRITE {{[*]*}} {{PV|T}}[[ADDRW:[0-9]*\.[XYZW]]]
-; EG-CHECK-NOT: LDS_WRITE {{[*]*}} T[[ADDRW]]
+; We would like to check the the lds writes are using different
+; addresses, but due to variations in the scheduler, we can't do
+; this consistently on evergreen GPUs.
+; EG-CHECK: LDS_WRITE
+; EG-CHECK: LDS_WRITE
 ; SI-CHECK: DS_WRITE_B32 0, {{v[0-9]*}}, v[[ADDRW:[0-9]*]]
 ; SI-CHECK-NOT: DS_WRITE_B32 0, {{v[0-9]*}}, v[[ADDRW]]