Adds cyclic critical path computation and heuristics, temporarily disabled.

Estimate the cyclic critical path within a single block loop. If the
acyclic critical path is longer, then the loop will exhaust OOO
resources after some number of iterations. If lag between the acyclic
critical path and cyclic critical path is longer the the time it takes
to issue those loop iterations, then aggressively schedule for
latency.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@189120 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/ScheduleDAGInstrs.h

@@ -197,6 +197,9 @@ namespace llvm {
     /// input.
     void buildSchedGraph(AliasAnalysis *AA, RegPressureTracker *RPTracker = 0);
 
+    /// Compute the cyclic critical path through the DAG.
+    unsigned computeCyclicCriticalPath();
+
     /// addSchedBarrierDeps - Add dependencies from instructions in the current
     /// list of instructions being scheduled to scheduling barrier. We want to
     /// make sure instructions which define registers that are either used by

lib/CodeGen/MachineScheduler.cpp

@@ -53,6 +53,9 @@ static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
 static bool ViewMISchedDAGs = false;
 #endif // NDEBUG
 
+static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden,
+  cl::desc("Enable cyclic critical path analysis."), cl::init(false));
+
 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
   cl::desc("Enable load clustering."), cl::init(true));
 
@@ -1207,16 +1210,21 @@ public:
   struct SchedRemainder {
     // Critical path through the DAG in expected latency.
     unsigned CriticalPath;
+    unsigned CyclicCritPath;
 
     // Scaled count of micro-ops left to schedule.
     unsigned RemIssueCount;
 
+    bool IsAcyclicLatencyLimited;
+
     // Unscheduled resources
     SmallVector<unsigned, 16> RemainingCounts;
 
     void reset() {
       CriticalPath = 0;
+      CyclicCritPath = 0;
       RemIssueCount = 0;
+      IsAcyclicLatencyLimited = false;
       RemainingCounts.clear();
     }
 
@@ -1434,6 +1442,8 @@ public:
   virtual void registerRoots();
 
 protected:
+  void checkAcyclicLatency();
+
   void tryCandidate(SchedCandidate &Cand,
                     SchedCandidate &TryCand,
                     SchedBoundary &Zone,
@@ -1547,8 +1557,32 @@ void ConvergingScheduler::releaseBottomNode(SUnit *SU) {
   Bot.releaseNode(SU, SU->BotReadyCycle);
 }
 
+void ConvergingScheduler::checkAcyclicLatency() {
+  if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath)
+    return;
+
+  unsigned BufferLimit =
+    SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor();
+  unsigned LatencyLag = Rem.CriticalPath - Rem.CyclicCritPath;
+  Rem.IsAcyclicLatencyLimited =
+    (LatencyLag * SchedModel->getLatencyFactor()) > BufferLimit;
+
+  DEBUG(dbgs() << "BufferLimit " << BufferLimit << "u / "
+        << Rem.RemIssueCount << "u = "
+        << (BufferLimit + Rem.RemIssueCount) / Rem.RemIssueCount << " iters. "
+        << "Latency = " << LatencyLag << "c = "
+        << LatencyLag * SchedModel->getLatencyFactor() << "u\n";
+        if (Rem.IsAcyclicLatencyLimited)
+          dbgs() << "  ACYCLIC LATENCY LIMIT\n");
+}
+
 void ConvergingScheduler::registerRoots() {
   Rem.CriticalPath = DAG->ExitSU.getDepth();
+
+  if (EnableCyclicPath) {
+    Rem.CyclicCritPath = DAG->computeCyclicCriticalPath();
+    checkAcyclicLatency();
+  }
   // Some roots may not feed into ExitSU. Check all of them in case.
   for (std::vector<SUnit*>::const_iterator
          I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) {
@@ -2096,6 +2130,32 @@ static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
   return 0;
 }
 
+static bool tryLatency(ConvergingScheduler::SchedCandidate &TryCand,
+                       ConvergingScheduler::SchedCandidate &Cand,
+                       ConvergingScheduler::SchedBoundary &Zone) {
+  if (Zone.isTop()) {
+    if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
+      if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
+                  TryCand, Cand, ConvergingScheduler::TopDepthReduce))
+        return true;
+    }
+    if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
+                   TryCand, Cand, ConvergingScheduler::TopPathReduce))
+      return true;
+  }
+  else {
+    if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
+      if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
+                  TryCand, Cand, ConvergingScheduler::BotHeightReduce))
+        return true;
+    }
+    if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
+                   TryCand, Cand, ConvergingScheduler::BotPathReduce))
+      return true;
+  }
+  return false;
+}
+
 /// Apply a set of heursitics to a new candidate. Heuristics are currently
 /// hierarchical. This may be more efficient than a graduated cost model because
 /// we don't need to evaluate all aspects of the model for each node in the
@@ -2135,6 +2195,10 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
                   RegExcess))
     return;
 
+  // For loops that are acyclic path limited, aggressively schedule for latency.
+  if (Rem.IsAcyclicLatencyLimited && tryLatency(TryCand, Cand, Zone))
+    return;
+
   // Avoid increasing the max critical pressure in the scheduled region.
   if (tryPressure(TryCand.RPDelta.CriticalMax, Cand.RPDelta.CriticalMax,
                   TryCand, Cand, RegCritical))
@@ -2174,27 +2238,10 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
     return;
 
   // Avoid serializing long latency dependence chains.
-  if (Cand.Policy.ReduceLatency) {
+  // For acyclic path limited loops, latency was already checked above.
-    if (Zone.isTop()) {
+  if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited
-      if (Cand.SU->getDepth() > Zone.getScheduledLatency()) {
+      && tryLatency(TryCand, Cand, Zone)) {
-        if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(),
+    return;
-                    TryCand, Cand, TopDepthReduce))
-          return;
-      }
-      if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(),
-                     TryCand, Cand, TopPathReduce))
-        return;
-    }
-    else {
-      if (Cand.SU->getHeight() > Zone.getScheduledLatency()) {
-        if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(),
-                    TryCand, Cand, BotHeightReduce))
-          return;
-      }
-      if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(),
-                     TryCand, Cand, BotPathReduce))
-        return;
-    }
   }
 
   // Prefer immediate defs/users of the last scheduled instruction. This is a

lib/CodeGen/ScheduleDAGInstrs.cpp

@@ -36,6 +36,8 @@
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetRegisterInfo.h"
 #include "llvm/Target/TargetSubtargetInfo.h"
+#include <queue>
+
 using namespace llvm;
 
 static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
@@ -979,6 +981,65 @@ void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
   PendingLoads.clear();
 }
 
+/// Compute the max cyclic critical path through the DAG. For loops that span
+/// basic blocks, MachineTraceMetrics should be used for this instead.
+unsigned ScheduleDAGInstrs::computeCyclicCriticalPath() {
+  // This only applies to single block loop.
+  if (!BB->isSuccessor(BB))
+    return 0;
+
+  unsigned MaxCyclicLatency = 0;
+  // Visit each live out vreg def to find def/use pairs that cross iterations.
+  for (SUnit::const_pred_iterator
+         PI = ExitSU.Preds.begin(), PE = ExitSU.Preds.end(); PI != PE; ++PI) {
+    MachineInstr *MI = PI->getSUnit()->getInstr();
+    for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
+      const MachineOperand &MO = MI->getOperand(i);
+      if (!MO.isReg() || !MO.isDef())
+        break;
+      unsigned Reg = MO.getReg();
+      if (!Reg || TRI->isPhysicalRegister(Reg))
+        continue;
+
+      const LiveInterval &LI = LIS->getInterval(Reg);
+      unsigned LiveOutHeight = PI->getSUnit()->getHeight();
+      unsigned LiveOutDepth = PI->getSUnit()->getDepth() + PI->getLatency();
+      // Visit all local users of the vreg def.
+      for (VReg2UseMap::iterator
+             UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) {
+        if (UI->SU == &ExitSU)
+          continue;
+
+        // Only consider uses of the phi.
+        LiveRangeQuery LRQ(LI, LIS->getInstructionIndex(UI->SU->getInstr()));
+        if (!LRQ.valueIn()->isPHIDef())
+          continue;
+
+        // Cheat a bit and assume that a path spanning two iterations is a
+        // cycle, which could overestimate in strange cases. This allows cyclic
+        // latency to be estimated as the minimum height or depth slack.
+        unsigned CyclicLatency = 0;
+        if (LiveOutDepth > UI->SU->getDepth())
+          CyclicLatency = LiveOutDepth - UI->SU->getDepth();
+        unsigned LiveInHeight = UI->SU->getHeight() + PI->getLatency();
+        if (LiveInHeight > LiveOutHeight) {
+          if (LiveInHeight - LiveOutHeight < CyclicLatency)
+            CyclicLatency = LiveInHeight - LiveOutHeight;
+        }
+        else
+          CyclicLatency = 0;
+        DEBUG(dbgs() << "Cyclic Path: SU(" << PI->getSUnit()->NodeNum
+              << ") -> SU(" << UI->SU->NodeNum << ") = "
+              << CyclicLatency << "\n");
+        if (CyclicLatency > MaxCyclicLatency)
+          MaxCyclicLatency = CyclicLatency;
+      }
+    }
+  }
+  DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "\n");
+  return MaxCyclicLatency;
+}
+
 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
   SU->getInstr()->dump();