Unbreak the JIT with SSE
[oota-llvm.git] / lib / Target / TargetSchedInfo.cpp
index d64652398ae7421eb8b7cb7e5220ecf127b4377b..a7ec1b545a374ab1777b63eb014af0c4b17e86a9 100644 (file)
@@ -1,14 +1,40 @@
 //===-- SchedInfo.cpp - Generic code to support target schedulers ----------==//
 //
+//                     The LLVM Compiler Infrastructure
+//
+// This file was developed by the LLVM research group and is distributed under
+// the University of Illinois Open Source License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
 // This file implements the generic part of a Scheduler description for a
 // target.  This functionality is defined in the llvm/Target/SchedInfo.h file.
 //
 //===----------------------------------------------------------------------===//
 
+#include "llvm/Config/alloca.h"
 #include "llvm/Target/TargetSchedInfo.h"
 #include "llvm/Target/TargetMachine.h"
+#include <algorithm>
+#include <iostream>
+using namespace llvm;
+
+resourceId_t llvm::CPUResource::nextId = 0;
+static std::vector<CPUResource*> *CPUResourceMap = 0;
 
-resourceId_t MachineResource::nextId = 0;
+CPUResource::CPUResource(const std::string& resourceName, int maxUsers)
+    : rname(resourceName), rid(nextId++), maxNumUsers(maxUsers) {
+  if(!CPUResourceMap)
+    CPUResourceMap = new std::vector<CPUResource*>;
+
+  //Put Resource in the map
+  CPUResourceMap->push_back(this);
+}
+
+///Get CPUResource if you only have the resource ID
+CPUResource* CPUResource::getCPUResource(resourceId_t id) {
+  return (*CPUResourceMap)[id];
+}
 
 // Check if fromRVec and toRVec have *any* common entries.
 // Assume the vectors are sorted in increasing order.
@@ -17,60 +43,56 @@ resourceId_t MachineResource::nextId = 0;
 //
 inline static bool
 RUConflict(const std::vector<resourceId_t>& fromRVec,
-          const std::vector<resourceId_t>& toRVec)
+           const std::vector<resourceId_t>& toRVec)
 {
-  
-  unsigned fN = fromRVec.size(), tN = toRVec.size(); 
+
+  unsigned fN = fromRVec.size(), tN = toRVec.size();
   unsigned fi = 0, ti = 0;
 
-  while (fi < fN && ti < tN)
-    {
-      if (fromRVec[fi] < toRVec[ti])
-       ++fi;
-      else if (toRVec[ti] < fromRVec[fi])
-       ++ti;
-      else
-       return true;
-    }
+  while (fi < fN && ti < tN) {
+    if (fromRVec[fi] < toRVec[ti])
+      ++fi;
+    else if (toRVec[ti] < fromRVec[fi])
+      ++ti;
+    else
+      return true;
+  }
   return false;
 }
 
 
-static cycles_t
-ComputeMinGap(const InstrRUsage &fromRU, 
-             const InstrRUsage &toRU)
+static CycleCount_t
+ComputeMinGap(const InstrRUsage &fromRU,
+              const InstrRUsage &toRU)
 {
-  cycles_t minGap = 0;
-  
+  CycleCount_t minGap = 0;
+
   if (fromRU.numBubbles > 0)
     minGap = fromRU.numBubbles;
-  
-  if (minGap < fromRU.numCycles)
-    {
-      // only need to check from cycle `minGap' onwards
-      for (cycles_t gap=minGap; gap <= fromRU.numCycles-1; gap++)
-       {
-         // check if instr. #2 can start executing `gap' cycles after #1
-         // by checking for resource conflicts in each overlapping cycle
-         cycles_t numOverlap =std::min(fromRU.numCycles - gap, toRU.numCycles);
-         for (cycles_t c = 0; c <= numOverlap-1; c++)
-           if (RUConflict(fromRU.resourcesByCycle[gap + c],
-                          toRU.resourcesByCycle[c]))
-             {
-               // conflict found so minGap must be more than `gap'
-               minGap = gap+1;
-               break;
-             }
-       }
+
+  if (minGap < fromRU.numCycles) {
+    // only need to check from cycle `minGap' onwards
+    for (CycleCount_t gap=minGap; gap <= fromRU.numCycles-1; gap++) {
+      // check if instr. #2 can start executing `gap' cycles after #1
+      // by checking for resource conflicts in each overlapping cycle
+      CycleCount_t numOverlap =std::min(fromRU.numCycles - gap, toRU.numCycles);
+      for (CycleCount_t c = 0; c <= numOverlap-1; c++)
+        if (RUConflict(fromRU.resourcesByCycle[gap + c],
+                       toRU.resourcesByCycle[c])) {
+          // conflict found so minGap must be more than `gap'
+          minGap = gap+1;
+          break;
+        }
     }
-  
+  }
+
   return minGap;
 }
 
 
 //---------------------------------------------------------------------------
 // class TargetSchedInfo
-//     Interface to machine description for instruction scheduling
+// Interface to machine description for instruction scheduling
 //---------------------------------------------------------------------------
 
 TargetSchedInfo::TargetSchedInfo(const TargetMachine&    tgt,
@@ -81,7 +103,7 @@ TargetSchedInfo::TargetSchedInfo(const TargetMachine&    tgt,
                                  unsigned NumUsageDeltas,
                                  unsigned NumIssueDeltas)
   : target(tgt),
-    numSchedClasses(NumSchedClasses), mii(tgt.getInstrInfo()),
+    numSchedClasses(NumSchedClasses), mii(tgt.getInstrInfo()),
     classRUsages(ClassRUsages), usageDeltas(UsageDeltas),
     issueDeltas(IssueDeltas), numUsageDeltas(NumUsageDeltas),
     numIssueDeltas(NumIssueDeltas)
@@ -90,13 +112,13 @@ TargetSchedInfo::TargetSchedInfo(const TargetMachine&    tgt,
 void
 TargetSchedInfo::initializeResources()
 {
-  assert(MAX_NUM_SLOTS >= (int)getMaxNumIssueTotal()
-        && "Insufficient slots for static data! Increase MAX_NUM_SLOTS");
-  
+  assert(MAX_NUM_SLOTS >= (int)getMaxNumIssueTotal() &&
+         "Insufficient slots for static data! Increase MAX_NUM_SLOTS");
+
   // First, compute common resource usage info for each class because
   // most instructions will probably behave the same as their class.
   // Cannot allocate a vector of InstrRUsage so new each one.
-  // 
+  //
   std::vector<InstrRUsage> instrRUForClasses;
   instrRUForClasses.resize(numSchedClasses);
   for (InstrSchedClass sc = 0; sc < numSchedClasses; sc++) {
@@ -104,7 +126,7 @@ TargetSchedInfo::initializeResources()
     instrRUForClasses[sc].setMaxSlots(getMaxNumIssueTotal());
     instrRUForClasses[sc].setTo(classRUsages[sc]);
   }
-  
+
   computeInstrResources(instrRUForClasses);
   computeIssueGaps(instrRUForClasses);
 }
@@ -112,25 +134,24 @@ TargetSchedInfo::initializeResources()
 
 void
 TargetSchedInfo::computeInstrResources(const std::vector<InstrRUsage>&
-                                       instrRUForClasses)
-{
-  int numOpCodes =  mii->getNumRealOpCodes();
+                                        instrRUForClasses) {
+  int numOpCodes =  mii->getNumOpcodes();
   instrRUsages.resize(numOpCodes);
-  
+
   // First get the resource usage information from the class resource usages.
   for (MachineOpCode op = 0; op < numOpCodes; ++op) {
     InstrSchedClass sc = getSchedClass(op);
     assert(sc < numSchedClasses);
     instrRUsages[op] = instrRUForClasses[sc];
   }
-  
+
   // Now, modify the resource usages as specified in the deltas.
   for (unsigned i = 0; i < numUsageDeltas; ++i) {
     MachineOpCode op = usageDeltas[i].opCode;
     assert(op < numOpCodes);
     instrRUsages[op].addUsageDelta(usageDeltas[i]);
   }
-  
+
   // Then modify the issue restrictions as specified in the deltas.
   for (unsigned i = 0; i < numIssueDeltas; ++i) {
     MachineOpCode op = issueDeltas[i].opCode;
@@ -142,27 +163,23 @@ TargetSchedInfo::computeInstrResources(const std::vector<InstrRUsage>&
 
 void
 TargetSchedInfo::computeIssueGaps(const std::vector<InstrRUsage>&
-                                  instrRUForClasses)
-{
-  int numOpCodes =  mii->getNumRealOpCodes();
+                                   instrRUForClasses) {
+  int numOpCodes =  mii->getNumOpcodes();
   issueGaps.resize(numOpCodes);
   conflictLists.resize(numOpCodes);
 
-  assert(numOpCodes < (1 << MAX_OPCODE_SIZE) - 1
-         && "numOpCodes invalid for implementation of class OpCodePair!");
-
   // First, compute issue gaps between pairs of classes based on common
   // resources usages for each class, because most instruction pairs will
   // usually behave the same as their class.
-  // 
-  int classPairGaps[numSchedClasses][numSchedClasses];
+  //
+  int* classPairGaps =
+    static_cast<int*>(alloca(sizeof(int) * numSchedClasses * numSchedClasses));
   for (InstrSchedClass fromSC=0; fromSC < numSchedClasses; fromSC++)
-    for (InstrSchedClass toSC=0; toSC < numSchedClasses; toSC++)
-      {
-       int classPairGap = ComputeMinGap(instrRUForClasses[fromSC],
-                                        instrRUForClasses[toSC]);
-       classPairGaps[fromSC][toSC] = classPairGap; 
-      }
+    for (InstrSchedClass toSC=0; toSC < numSchedClasses; toSC++) {
+      int classPairGap = ComputeMinGap(instrRUForClasses[fromSC],
+                                       instrRUForClasses[toSC]);
+      classPairGaps[fromSC*numSchedClasses + toSC] = classPairGap;
+    }
 
   // Now, for each pair of instructions, use the class pair gap if both
   // instructions have identical resource usage as their respective classes.
@@ -171,84 +188,79 @@ TargetSchedInfo::computeIssueGaps(const std::vector<InstrRUsage>&
   longestIssueConflict = 0;
 
   for (MachineOpCode fromOp=0; fromOp < numOpCodes; fromOp++)
-    for (MachineOpCode toOp=0; toOp < numOpCodes; toOp++)
-      {
-        int instrPairGap = 
-          (instrRUsages[fromOp].sameAsClass && instrRUsages[toOp].sameAsClass)
-          ? classPairGaps[getSchedClass(fromOp)][getSchedClass(toOp)]
-          : ComputeMinGap(instrRUsages[fromOp], instrRUsages[toOp]);
-
-        if (instrPairGap > 0)
-          {
-            this->setGap(instrPairGap, fromOp, toOp);
-            conflictLists[fromOp].push_back(toOp);
-            longestIssueConflict=std::max(longestIssueConflict, instrPairGap);
-          }
+    for (MachineOpCode toOp=0; toOp < numOpCodes; toOp++) {
+      int instrPairGap =
+        (instrRUsages[fromOp].sameAsClass && instrRUsages[toOp].sameAsClass)
+        ? classPairGaps[getSchedClass(fromOp)*numSchedClasses + getSchedClass(toOp)]
+        : ComputeMinGap(instrRUsages[fromOp], instrRUsages[toOp]);
+
+      if (instrPairGap > 0) {
+        this->setGap(instrPairGap, fromOp, toOp);
+        conflictLists[fromOp].push_back(toOp);
+        longestIssueConflict=std::max(longestIssueConflict, instrPairGap);
       }
+    }
 }
 
 
 void InstrRUsage::setTo(const InstrClassRUsage& classRU) {
-  sameAsClass  = true;
+  sameAsClass   = true;
   isSingleIssue = classRU.isSingleIssue;
-  breaksGroup   = classRU.breaksGroup; 
+  breaksGroup   = classRU.breaksGroup;
   numBubbles    = classRU.numBubbles;
-  
-  for (unsigned i=0; i < classRU.numSlots; i++)
-    {
-      unsigned slot = classRU.feasibleSlots[i];
-      assert(slot < feasibleSlots.size() && "Invalid slot specified!");
-      this->feasibleSlots[slot] = true;
-    }
-  
+
+  for (unsigned i=0; i < classRU.numSlots; i++) {
+    unsigned slot = classRU.feasibleSlots[i];
+    assert(slot < feasibleSlots.size() && "Invalid slot specified!");
+    this->feasibleSlots[slot] = true;
+  }
+
   numCycles   = classRU.totCycles;
   resourcesByCycle.resize(this->numCycles);
-  
+
   for (unsigned i=0; i < classRU.numRUEntries; i++)
     for (unsigned c=classRU.V[i].startCycle, NC = c + classRU.V[i].numCycles;
-        c < NC; c++)
+         c < NC; c++)
       this->resourcesByCycle[c].push_back(classRU.V[i].resourceId);
-  
-  // Sort each resource usage vector by resourceId_t to speed up conflict checking
+
+  // Sort each resource usage vector by resourceId_t to speed up conflict
+  // checking
   for (unsigned i=0; i < this->resourcesByCycle.size(); i++)
-    sort(resourcesByCycle[i].begin(), resourcesByCycle[i].end());
-  
+    std::sort(resourcesByCycle[i].begin(), resourcesByCycle[i].end());
 }
 
 // Add the extra resource usage requirements specified in the delta.
 // Note that a negative value of `numCycles' means one entry for that
 // resource should be deleted for each cycle.
-// 
+//
 void InstrRUsage::addUsageDelta(const InstrRUsageDelta &delta) {
   int NC = delta.numCycles;
   sameAsClass = false;
-  
+
   // resize the resources vector if more cycles are specified
   unsigned maxCycles = this->numCycles;
   maxCycles = std::max(maxCycles, delta.startCycle + abs(NC) - 1);
-  if (maxCycles > this->numCycles)
-    {
-      this->resourcesByCycle.resize(maxCycles);
-      this->numCycles = maxCycles;
-    }
-    
+  if (maxCycles > this->numCycles) {
+    this->resourcesByCycle.resize(maxCycles);
+    this->numCycles = maxCycles;
+  }
+
   if (NC >= 0)
     for (unsigned c=delta.startCycle, last=c+NC-1; c <= last; c++)
       this->resourcesByCycle[c].push_back(delta.resourceId);
   else
     // Remove the resource from all NC cycles.
-    for (unsigned c=delta.startCycle, last=(c-NC)-1; c <= last; c++)
-      {
-       // Look for the resource backwards so we remove the last entry
-       // for that resource in each cycle.
-       std::vector<resourceId_t>& rvec = this->resourcesByCycle[c];
-       int r;
-       for (r = (int) rvec.size(); r >= 0; r--)
-         if (rvec[r] == delta.resourceId)
-           {// found last entry for the resource
-             rvec.erase(rvec.begin() + r);
-             break;
-           }
-       assert(r >= 0 && "Resource to remove was unused in cycle c!");
-      }
+    for (unsigned c=delta.startCycle, last=(c-NC)-1; c <= last; c++) {
+      // Look for the resource backwards so we remove the last entry
+      // for that resource in each cycle.
+      std::vector<resourceId_t>& rvec = this->resourcesByCycle[c];
+      int r;
+      for (r = rvec.size() - 1; r >= 0; r--)
+        if (rvec[r] == delta.resourceId) {
+          // found last entry for the resource
+          rvec.erase(rvec.begin() + r);
+          break;
+        }
+      assert(r >= 0 && "Resource to remove was unused in cycle c!");
+    }
 }