Provide operand indices to adjustSchedDependency

This allows targets to know exactly which operands are contributing to the dependency, which is required for targets with per-operand scheduling models. Differential Revision: https://reviews.llvm.org/D77135
llvm · Apr 17, 2020 · c819ef9 · c819ef9
1 parent fa7f328
commit c819ef9
Show file tree

Hide file tree

Showing 8 changed files with 30 additions and 22 deletions.
diff --git a/llvm/include/llvm/CodeGen/TargetSubtargetInfo.h b/llvm/include/llvm/CodeGen/TargetSubtargetInfo.h
@@ -224,9 +224,13 @@ class TargetSubtargetInfo : public MCSubtargetInfo {
   virtual void overrideSchedPolicy(MachineSchedPolicy &Policy,
                                    unsigned NumRegionInstrs) const {}
 
-  // Perform target specific adjustments to the latency of a schedule
+  // Perform target-specific adjustments to the latency of a schedule
   // dependency.
-  virtual void adjustSchedDependency(SUnit *def, SUnit *use, SDep &dep) const {}
+  // If a pair of operands is associated with the schedule dependency, DefOpIdx
+  // and UseOpIdx are the indices of the operands in Def and Use, respectively.
+  // Otherwise, either may be -1.
+  virtual void adjustSchedDependency(SUnit *Def, int DefOpIdx, SUnit *Use,
+                                     int UseOpIdx, SDep &Dep) const {}
 
   // For use with PostRAScheduling: get the anti-dependence breaking that should
   // be performed before post-RA scheduling.

diff --git a/llvm/lib/CodeGen/MachinePipeliner.cpp b/llvm/lib/CodeGen/MachinePipeliner.cpp
@@ -809,7 +809,7 @@ void SwingSchedulerDAG::updatePhiDependences() {
           if (!MI->isPHI()) {
             SDep Dep(SU, SDep::Data, Reg);
             Dep.setLatency(0);
-            ST.adjustSchedDependency(SU, &I, Dep);
+            ST.adjustSchedDependency(SU, 0, &I, MI->getOperandNo(MOI), Dep);
             I.addPred(Dep);
           } else {
             HasPhiUse = Reg;

diff --git a/llvm/lib/CodeGen/ScheduleDAGInstrs.cpp b/llvm/lib/CodeGen/ScheduleDAGInstrs.cpp
@@ -269,13 +269,13 @@ void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
       if (!ImplicitPseudoDef && !ImplicitPseudoUse) {
         Dep.setLatency(SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
                                                         RegUse, UseOp));
-        ST.adjustSchedDependency(SU, UseSU, Dep);
+        ST.adjustSchedDependency(SU, OperIdx, UseSU, UseOp, Dep);
       } else {
         Dep.setLatency(0);
         // FIXME: We could always let target to adjustSchedDependency(), and
         // remove this condition, but that currently asserts in Hexagon BE.
         if (SU->getInstr()->isBundle() || (RegUse && RegUse->isBundle()))
-          ST.adjustSchedDependency(SU, UseSU, Dep);
+          ST.adjustSchedDependency(SU, OperIdx, UseSU, UseOp, Dep);
       }
 
       UseSU->addPred(Dep);
@@ -444,7 +444,7 @@ void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
         SDep Dep(SU, SDep::Data, Reg);
         Dep.setLatency(SchedModel.computeOperandLatency(MI, OperIdx, Use,
                                                         I->OperandIndex));
-        ST.adjustSchedDependency(SU, UseSU, Dep);
+        ST.adjustSchedDependency(SU, OperIdx, UseSU, I->OperandIndex, Dep);
         UseSU->addPred(Dep);
       }
 

diff --git a/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp b/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp
@@ -474,6 +474,7 @@ void ScheduleDAGSDNodes::AddSchedEdges() {
 
       for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
         SDNode *OpN = N->getOperand(i).getNode();
+        unsigned DefIdx = N->getOperand(i).getResNo();
         if (isPassiveNode(OpN)) continue;   // Not scheduled.
         SUnit *OpSU = &SUnits[OpN->getNodeId()];
         assert(OpSU && "Node has no SUnit!");
@@ -508,7 +509,7 @@ void ScheduleDAGSDNodes::AddSchedEdges() {
         Dep.setLatency(OpLatency);
         if (!isChain && !UnitLatencies) {
           computeOperandLatency(OpN, N, i, Dep);
-          ST.adjustSchedDependency(OpSU, SU, Dep);
+          ST.adjustSchedDependency(OpSU, DefIdx, SU, i, Dep);
         }
 
         if (!SU->addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -722,20 +722,20 @@ unsigned GCNSubtarget::getMaxNumVGPRs(const MachineFunction &MF) const {
   return MaxNumVGPRs;
 }
 
-void GCNSubtarget::adjustSchedDependency(SUnit *Src, SUnit *Dst,
-                                         SDep &Dep) const {
+void GCNSubtarget::adjustSchedDependency(SUnit *Def, int DefOpIdx, SUnit *Use,
+                                         int UseOpIdx, SDep &Dep) const {
   if (Dep.getKind() != SDep::Kind::Data || !Dep.getReg() ||
-      !Src->isInstr() || !Dst->isInstr())
+      !Def->isInstr() || !Use->isInstr())
     return;
 
-  MachineInstr *SrcI = Src->getInstr();
-  MachineInstr *DstI = Dst->getInstr();
+  MachineInstr *DefI = Def->getInstr();
+  MachineInstr *UseI = Use->getInstr();
 
-  if (SrcI->isBundle()) {
+  if (DefI->isBundle()) {
     const SIRegisterInfo *TRI = getRegisterInfo();
     auto Reg = Dep.getReg();
-    MachineBasicBlock::const_instr_iterator I(SrcI->getIterator());
-    MachineBasicBlock::const_instr_iterator E(SrcI->getParent()->instr_end());
+    MachineBasicBlock::const_instr_iterator I(DefI->getIterator());
+    MachineBasicBlock::const_instr_iterator E(DefI->getParent()->instr_end());
     unsigned Lat = 0;
     for (++I; I != E && I->isBundledWithPred(); ++I) {
       if (I->modifiesRegister(Reg, TRI))
@@ -744,12 +744,12 @@ void GCNSubtarget::adjustSchedDependency(SUnit *Src, SUnit *Dst,
         --Lat;
     }
     Dep.setLatency(Lat);
-  } else if (DstI->isBundle()) {
+  } else if (UseI->isBundle()) {
     const SIRegisterInfo *TRI = getRegisterInfo();
     auto Reg = Dep.getReg();
-    MachineBasicBlock::const_instr_iterator I(DstI->getIterator());
-    MachineBasicBlock::const_instr_iterator E(DstI->getParent()->instr_end());
-    unsigned Lat = InstrInfo.getInstrLatency(getInstrItineraryData(), *SrcI);
+    MachineBasicBlock::const_instr_iterator I(UseI->getIterator());
+    MachineBasicBlock::const_instr_iterator E(UseI->getParent()->instr_end());
+    unsigned Lat = InstrInfo.getInstrLatency(getInstrItineraryData(), *DefI);
     for (++I; I != E && I->isBundledWithPred() && Lat; ++I) {
       if (I->readsRegister(Reg, TRI))
         break;

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -1193,7 +1193,8 @@ class GCNSubtarget : public AMDGPUGenSubtargetInfo,
     return AMDGPU::IsaInfo::getMinWavesPerEU(this);
   }
 
-  void adjustSchedDependency(SUnit *Src, SUnit *Dst, SDep &Dep) const override;
+  void adjustSchedDependency(SUnit *Def, int DefOpIdx, SUnit *Use, int UseOpIdx,
+                             SDep &Dep) const override;
 };
 
 class R600Subtarget final : public R600GenSubtargetInfo,

diff --git a/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp b/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp
@@ -315,7 +315,8 @@ bool HexagonSubtarget::useAA() const {
 
 /// Perform target specific adjustments to the latency of a schedule
 /// dependency.
-void HexagonSubtarget::adjustSchedDependency(SUnit *Src, SUnit *Dst,
+void HexagonSubtarget::adjustSchedDependency(SUnit *Src, int SrcOpIdx,
+                                             SUnit *Dst, int DstOpIdx,
                                              SDep &Dep) const {
   MachineInstr *SrcInst = Src->getInstr();
   MachineInstr *DstInst = Dst->getInstr();

diff --git a/llvm/lib/Target/Hexagon/HexagonSubtarget.h b/llvm/lib/Target/Hexagon/HexagonSubtarget.h
@@ -258,7 +258,8 @@ class HexagonSubtarget : public HexagonGenSubtargetInfo {
 
   /// Perform target specific adjustments to the latency of a schedule
   /// dependency.
-  void adjustSchedDependency(SUnit *def, SUnit *use, SDep& dep) const override;
+  void adjustSchedDependency(SUnit *Def, int DefOpIdx, SUnit *Use, int UseOpIdx,
+                             SDep &Dep) const override;
 
   unsigned getVectorLength() const {
     assert(useHVXOps());