Machine Model: Add MicroOpBufferSize and resource BufferSize.

Replace the ill-defined MinLatency and ILPWindow properties with
with straightforward buffer sizes:
MCSchedMode::MicroOpBufferSize
MCProcResourceDesc::BufferSize

These can be used to more precisely model instruction execution if desired.

Disabled some misched tests temporarily. They'll be reenabled in a few commits.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@184032 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 47 insertions, 102 deletions

diff --git a/lib/CodeGen/MachineScheduler.cpp b/lib/CodeGen/MachineScheduler.cpp
index 3840473a9a..c87a1be720 100644
--- a/lib/CodeGen/MachineScheduler.cpp
+++ b/lib/CodeGen/MachineScheduler.cpp
@@ -1257,8 +1257,9 @@ public:
     unsigned ExpectedCount;
 
 #ifndef NDEBUG
-    // Remember the greatest min operand latency.
-    unsigned MaxMinLatency;
+    // Remember the greatest operand latency as an upper bound on the number of
+    // times we should retry the pending queue because of a hazard.
+    unsigned MaxObservedLatency;
 #endif
 
     void reset() {
@@ -1281,7 +1282,7 @@ public:
       IsResourceLimited = false;
       ExpectedCount = 0;
 #ifndef NDEBUG
-      MaxMinLatency = 0;
+      MaxObservedLatency = 0;
 #endif
       // Reserve a zero-count for invalid CritResIdx.
       ResourceCounts.resize(1);
@@ -1466,13 +1467,15 @@ void ConvergingScheduler::releaseTopNode(SUnit *SU) {
 
   for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
        I != E; ++I) {
+    if (I->isWeak())
+      continue;
     unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle;
-    unsigned MinLatency = I->getMinLatency();
+    unsigned Latency = I->getLatency();
 #ifndef NDEBUG
-    Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency);
+    Top.MaxObservedLatency = std::max(Latency, Top.MaxObservedLatency);
 #endif
-    if (SU->TopReadyCycle < PredReadyCycle + MinLatency)
-      SU->TopReadyCycle = PredReadyCycle + MinLatency;
+    if (SU->TopReadyCycle < PredReadyCycle + Latency)
+      SU->TopReadyCycle = PredReadyCycle + Latency;
   }
   Top.releaseNode(SU, SU->TopReadyCycle);
 }
@@ -1488,12 +1491,12 @@ void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
     if (I->isWeak())
       continue;
     unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle;
-    unsigned MinLatency = I->getMinLatency();
+    unsigned Latency = I->getLatency();
 #ifndef NDEBUG
-    Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency);
+    Bot.MaxObservedLatency = std::max(Latency, Bot.MaxObservedLatency);
 #endif
-    if (SU->BotReadyCycle < SuccReadyCycle + MinLatency)
-      SU->BotReadyCycle = SuccReadyCycle + MinLatency;
+    if (SU->BotReadyCycle < SuccReadyCycle + Latency)
+      SU->BotReadyCycle = SuccReadyCycle + Latency;
   }
   Bot.releaseNode(SU, SU->BotReadyCycle);
 }
@@ -1558,7 +1561,7 @@ void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
     if (L > RemLatency)
       RemLatency = L;
   }
-  unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
+  unsigned CriticalPathLimit = Rem->CriticalPath;
   DEBUG(dbgs() << "  " << Available.getName()
         << " ExpectedLatency " << ExpectedLatency
         << " CP Limit " << CriticalPathLimit << '\n');
@@ -1751,7 +1754,7 @@ SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() {
     }
   }
   for (unsigned i = 0; Available.empty(); ++i) {
-    assert(i <= (HazardRec->getMaxLookAhead() + MaxMinLatency) &&
+    assert(i <= (HazardRec->getMaxLookAhead() + MaxObservedLatency) &&
            "permanent hazard"); (void)i;
     bumpCycle();
     releasePending();
diff --git a/lib/CodeGen/MachineTraceMetrics.cpp b/lib/CodeGen/MachineTraceMetrics.cpp
index 00f702c846..6aa3f67962 100644
--- a/lib/CodeGen/MachineTraceMetrics.cpp
+++ b/lib/CodeGen/MachineTraceMetrics.cpp
@@ -853,8 +853,7 @@ computeInstrDepths(const MachineBasicBlock *MBB) {
         // Add latency if DefMI is a real instruction. Transients get latency 0.
         if (!Dep.DefMI->isTransient())
           DepCycle += MTM.SchedModel
-            .computeOperandLatency(Dep.DefMI, Dep.DefOp, UseMI, Dep.UseOp,
-                                   /* FindMin = */ false);
+            .computeOperandLatency(Dep.DefMI, Dep.DefOp, UseMI, Dep.UseOp);
         Cycle = std::max(Cycle, DepCycle);
       }
       // Remember the instruction depth.
@@ -902,8 +901,7 @@ static unsigned updatePhysDepsUpwards(const MachineInstr *MI, unsigned Height,
         // We may not know the UseMI of this dependency, if it came from the
         // live-in list. SchedModel can handle a NULL UseMI.
         DepHeight += SchedModel
-          .computeOperandLatency(MI, MO.getOperandNo(), I->MI, I->Op,
-                                 /* FindMin = */ false);
+          .computeOperandLatency(MI, MO.getOperandNo(), I->MI, I->Op);
       }
       Height = std::max(Height, DepHeight);
       // This regunit is dead above MI.
@@ -941,7 +939,7 @@ static bool pushDepHeight(const DataDep &Dep,
   // Adjust height by Dep.DefMI latency.
   if (!Dep.DefMI->isTransient())
     UseHeight += SchedModel.computeOperandLatency(Dep.DefMI, Dep.DefOp,
-                                                  UseMI, Dep.UseOp, false);
+                                                  UseMI, Dep.UseOp);
 
   // Update Heights[DefMI] to be the maximum height seen.
   MIHeightMap::iterator I;
@@ -1171,7 +1169,7 @@ MachineTraceMetrics::Trace::getPHIDepth(const MachineInstr *PHI) const {
   // Add latency if DefMI is a real instruction. Transients get latency 0.
   if (!Dep.DefMI->isTransient())
     DepCycle += TE.MTM.SchedModel
-      .computeOperandLatency(Dep.DefMI, Dep.DefOp, PHI, Dep.UseOp, false);
+      .computeOperandLatency(Dep.DefMI, Dep.DefOp, PHI, Dep.UseOp);
   return DepCycle;
 }
 
diff --git a/lib/CodeGen/ScheduleDAGInstrs.cpp b/lib/CodeGen/ScheduleDAGInstrs.cpp
index e4da6a41ee..aaf5c88a83 100644
--- a/lib/CodeGen/ScheduleDAGInstrs.cpp
+++ b/lib/CodeGen/ScheduleDAGInstrs.cpp
@@ -267,13 +267,10 @@ void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
         SU->hasPhysRegDefs = true;
         Dep = SDep(SU, SDep::Data, *Alias);
         RegUse = UseSU->getInstr();
-        Dep.setMinLatency(
-          SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
-                                           RegUse, UseOp, /*FindMin=*/true));
       }
       Dep.setLatency(
-        SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
-                                         RegUse, UseOp, /*FindMin=*/false));
+        SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
+                                         UseOp));
 
       ST.adjustSchedDependency(SU, UseSU, Dep);
       UseSU->addPred(Dep);
@@ -310,10 +307,8 @@ void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
           DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
         else {
           SDep Dep(SU, Kind, /*Reg=*/*Alias);
-          unsigned OutLatency =
-            SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
-          Dep.setMinLatency(OutLatency);
-          Dep.setLatency(OutLatency);
+          Dep.setLatency(
+            SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
           DefSU->addPred(Dep);
         }
       }
@@ -389,10 +384,8 @@ void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
     SUnit *DefSU = DefI->SU;
     if (DefSU != SU && DefSU != &ExitSU) {
       SDep Dep(SU, SDep::Output, Reg);
-      unsigned OutLatency =
-        SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr());
-      Dep.setMinLatency(OutLatency);
-      Dep.setLatency(OutLatency);
+      Dep.setLatency(
+        SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
       DefSU->addPred(Dep);
     }
     DefI->SU = SU;
@@ -427,10 +420,7 @@ void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
       // Adjust the dependence latency using operand def/use information, then
       // allow the target to perform its own adjustments.
       int DefOp = Def->findRegisterDefOperandIdx(Reg);
-      dep.setLatency(
-        SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, false));
-      dep.setMinLatency(
-        SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, true));
+      dep.setLatency(SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx));
 
       const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
       ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
diff --git a/lib/CodeGen/TargetInstrInfo.cpp b/lib/CodeGen/TargetInstrInfo.cpp
index 20eb918793..bb8bd426a9 100644
--- a/lib/CodeGen/TargetInstrInfo.cpp
+++ b/lib/CodeGen/TargetInstrInfo.cpp
@@ -668,27 +668,13 @@ getOperandLatency(const InstrItineraryData *ItinData,
 /// lookup, do so. Otherwise return -1.
 int TargetInstrInfo::computeDefOperandLatency(
   const InstrItineraryData *ItinData,
-  const MachineInstr *DefMI, bool FindMin) const {
+  const MachineInstr *DefMI) const {
 
   // Let the target hook getInstrLatency handle missing itineraries.
   if (!ItinData)
     return getInstrLatency(ItinData, DefMI);
 
-  // Return a latency based on the itinerary properties and defining instruction
-  // if possible. Some common subtargets don't require per-operand latency,
-  // especially for minimum latencies.
-  if (FindMin) {
-    // If MinLatency is valid, call getInstrLatency. This uses Stage latency if
-    // it exists before defaulting to MinLatency.
-    if (ItinData->SchedModel->MinLatency >= 0)
-      return getInstrLatency(ItinData, DefMI);
-
-    // If MinLatency is invalid, OperandLatency is interpreted as MinLatency.
-    // For empty itineraries, short-cirtuit the check and default to one cycle.
-    if (ItinData->isEmpty())
-      return 1;
-  }
-  else if(ItinData->isEmpty())
+  if(ItinData->isEmpty())
     return defaultDefLatency(ItinData->SchedModel, DefMI);
 
   // ...operand lookup required
@@ -709,10 +695,9 @@ int TargetInstrInfo::computeDefOperandLatency(
 unsigned TargetInstrInfo::
 computeOperandLatency(const InstrItineraryData *ItinData,
                       const MachineInstr *DefMI, unsigned DefIdx,
-                      const MachineInstr *UseMI, unsigned UseIdx,
-                      bool FindMin) const {
+                      const MachineInstr *UseMI, unsigned UseIdx) const {
 
-  int DefLatency = computeDefOperandLatency(ItinData, DefMI, FindMin);
+  int DefLatency = computeDefOperandLatency(ItinData, DefMI);
   if (DefLatency >= 0)
     return DefLatency;
 
@@ -732,8 +717,7 @@ computeOperandLatency(const InstrItineraryData *ItinData,
   unsigned InstrLatency = getInstrLatency(ItinData, DefMI);
 
   // Expected latency is the max of the stage latency and itinerary props.
-  if (!FindMin)
-    InstrLatency = std::max(InstrLatency,
-                            defaultDefLatency(ItinData->SchedModel, DefMI));
+  InstrLatency = std::max(InstrLatency,
+                          defaultDefLatency(ItinData->SchedModel, DefMI));
   return InstrLatency;
 }
diff --git a/lib/CodeGen/TargetSchedule.cpp b/lib/CodeGen/TargetSchedule.cpp
index 1bf14dbcef..53cd11c5da 100644
--- a/lib/CodeGen/TargetSchedule.cpp
+++ b/lib/CodeGen/TargetSchedule.cpp
@@ -93,33 +93,10 @@ unsigned TargetSchedModel::getNumMicroOps(const MachineInstr *MI,
 // effectively means infinite latency. Since users of the TargetSchedule API
 // don't know how to handle this, we convert it to a very large latency that is
 // easy to distinguish when debugging the DAG but won't induce overflow.
-static unsigned convertLatency(int Cycles) {
+static unsigned capLatency(int Cycles) {
   return Cycles >= 0 ? Cycles : 1000;
 }
 
-/// If we can determine the operand latency from the def only, without machine
-/// model or itinerary lookup, do so. Otherwise return -1.
-int TargetSchedModel::getDefLatency(const MachineInstr *DefMI,
-                                    bool FindMin) const {
-
-  // Return a latency based on the itinerary properties and defining instruction
-  // if possible. Some common subtargets don't require per-operand latency,
-  // especially for minimum latencies.
-  if (FindMin) {
-    // If MinLatency is invalid, then use the itinerary for MinLatency. If no
-    // itinerary exists either, then use single cycle latency.
-    if (SchedModel.MinLatency < 0 && !hasInstrItineraries()) {
-      return 1;
-    }
-    return SchedModel.MinLatency;
-  }
-  else if (!hasInstrSchedModel() && !hasInstrItineraries()) {
-    return TII->defaultDefLatency(&SchedModel, DefMI);
-  }
-  // ...operand lookup required
-  return -1;
-}
-
 /// Return the MCSchedClassDesc for this instruction. Some SchedClasses require
 /// evaluation of predicates that depend on instruction operands or flags.
 const MCSchedClassDesc *TargetSchedModel::
@@ -177,18 +154,16 @@ static unsigned findUseIdx(const MachineInstr *MI, unsigned UseOperIdx) {
 // Top-level API for clients that know the operand indices.
 unsigned TargetSchedModel::computeOperandLatency(
   const MachineInstr *DefMI, unsigned DefOperIdx,
-  const MachineInstr *UseMI, unsigned UseOperIdx,
-  bool FindMin) const {
+  const MachineInstr *UseMI, unsigned UseOperIdx) const {
 
-  int DefLatency = getDefLatency(DefMI, FindMin);
-  if (DefLatency >= 0)
-    return DefLatency;
+  if (!hasInstrSchedModel() && !hasInstrItineraries())
+    return TII->defaultDefLatency(&SchedModel, DefMI);
 
   if (hasInstrItineraries()) {
     int OperLatency = 0;
     if (UseMI) {
-      OperLatency =
-        TII->getOperandLatency(&InstrItins, DefMI, DefOperIdx, UseMI, UseOperIdx);
+      OperLatency = TII->getOperandLatency(&InstrItins, DefMI, DefOperIdx,
+                                           UseMI, UseOperIdx);
     }
     else {
       unsigned DefClass = DefMI->getDesc().getSchedClass();
@@ -205,13 +180,11 @@ unsigned TargetSchedModel::computeOperandLatency(
     // hook to allow subtargets to specialize latency. This hook is only
     // applicable to the InstrItins model. InstrSchedModel should model all
     // special cases without TII hooks.
-    if (!FindMin)
-      InstrLatency = std::max(InstrLatency,
-                              TII->defaultDefLatency(&SchedModel, DefMI));
+    InstrLatency = std::max(InstrLatency,
+                            TII->defaultDefLatency(&SchedModel, DefMI));
     return InstrLatency;
   }
-  assert(!FindMin && hasInstrSchedModel() &&
-         "Expected a SchedModel for this cpu");
+  // hasInstrSchedModel()
   const MCSchedClassDesc *SCDesc = resolveSchedClass(DefMI);
   unsigned DefIdx = findDefIdx(DefMI, DefOperIdx);
   if (DefIdx < SCDesc->NumWriteLatencyEntries) {
@@ -219,7 +192,7 @@ unsigned TargetSchedModel::computeOperandLatency(
     const MCWriteLatencyEntry *WLEntry =
       STI->getWriteLatencyEntry(SCDesc, DefIdx);
     unsigned WriteID = WLEntry->WriteResourceID;
-    unsigned Latency = convertLatency(WLEntry->Cycles);
+    unsigned Latency = capLatency(WLEntry->Cycles);
     if (!UseMI)
       return Latency;
 
@@ -263,7 +236,7 @@ unsigned TargetSchedModel::computeInstrLatency(const MachineInstr *MI) const {
         // Lookup the definition's write latency in SubtargetInfo.
         const MCWriteLatencyEntry *WLEntry =
           STI->getWriteLatencyEntry(SCDesc, DefIdx);
-        Latency = std::max(Latency, convertLatency(WLEntry->Cycles));
+        Latency = std::max(Latency, capLatency(WLEntry->Cycles));
       }
       return Latency;
     }
@@ -274,13 +247,10 @@ unsigned TargetSchedModel::computeInstrLatency(const MachineInstr *MI) const {
 unsigned TargetSchedModel::
 computeOutputLatency(const MachineInstr *DefMI, unsigned DefOperIdx,
                      const MachineInstr *DepMI) const {
-  // MinLatency == -1 is for in-order processors that always have unit
-  // MinLatency. MinLatency > 0 is for in-order processors with varying min
-  // latencies, but since this is not a RAW dep, we always use unit latency.
-  if (SchedModel.MinLatency != 0)
+  if (SchedModel.MicroOpBufferSize <= 1)
     return 1;
 
-  // MinLatency == 0 indicates an out-of-order processor that can dispatch
+  // MicroOpBufferSize > 1 indicates an out-of-order processor that can dispatch
   // WAW dependencies in the same cycle.
 
   // Treat predication as a data dependency for out-of-order cpus. In-order
@@ -302,7 +272,7 @@ computeOutputLatency(const MachineInstr *DefMI, unsigned DefOperIdx,
     if (SCDesc->isValid()) {
       for (const MCWriteProcResEntry *PRI = STI->getWriteProcResBegin(SCDesc),
              *PRE = STI->getWriteProcResEnd(SCDesc); PRI != PRE; ++PRI) {
-        if (!SchedModel.getProcResource(PRI->ProcResourceIdx)->IsBuffered)
+        if (!SchedModel.getProcResource(PRI->ProcResourceIdx)->BufferSize)
           return 1;
       }
     }