Fix the issues originally addressed in r54070. After thinking about it some more...
[oota-llvm.git] / lib / CodeGen / StrongPHIElimination.cpp
index 0c785367168c270f338656c9b69607749cd4ec9e..399e61541843c7490790283c54b8130e732f64ea 100644 (file)
@@ -27,6 +27,7 @@
 #include "llvm/CodeGen/MachineInstr.h"
 #include "llvm/CodeGen/MachineLoopInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/RegisterCoalescer.h"
 #include "llvm/Target/TargetInstrInfo.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/ADT/DepthFirstIterator.h"
@@ -34,7 +35,6 @@
 #include "llvm/Support/Compiler.h"
 using namespace llvm;
 
-
 namespace {
   struct VISIBILITY_HIDDEN StrongPHIElimination : public MachineFunctionPass {
     static char ID; // Pass identification, replacement for typeid
@@ -52,9 +52,10 @@ namespace {
     // used as operands to another another PHI node
     std::set<unsigned> UsedByAnother;
     
-    // RenameSets are the sets of operands (and their VNInfo IDs) to a PHI
-    // (the defining instruction of the key) that can be renamed without copies.
-    std::map<unsigned, std::map<unsigned, unsigned> > RenameSets;
+    // RenameSets are the is a map from a PHI-defined register
+    // to the input registers to be coalesced along with the 
+    // predecessor block for those input registers.
+    std::map<unsigned, std::map<unsigned, MachineBasicBlock*> > RenameSets;
     
     // PhiValueNumber holds the ID numbers of the VNs for each phi that we're
     // eliminating, indexed by the register defined by that phi.
@@ -74,6 +75,7 @@ namespace {
       
       // TODO: Actually make this true.
       AU.addPreserved<LiveIntervals>();
+      AU.addPreserved<RegisterCoalescer>();
       MachineFunctionPass::getAnalysisUsage(AU);
     }
     
@@ -129,24 +131,27 @@ namespace {
     void computeDFS(MachineFunction& MF);
     void processBlock(MachineBasicBlock* MBB);
     
-    std::vector<DomForestNode*> computeDomForest(std::map<unsigned, unsigned>& instrs,
+    std::vector<DomForestNode*> computeDomForest(
+                           std::map<unsigned, MachineBasicBlock*>& instrs,
                                                  MachineRegisterInfo& MRI);
     void processPHIUnion(MachineInstr* Inst,
-                         std::map<unsigned, unsigned>& PHIUnion,
+                         std::map<unsigned, MachineBasicBlock*>& PHIUnion,
                          std::vector<StrongPHIElimination::DomForestNode*>& DF,
                          std::vector<std::pair<unsigned, unsigned> >& locals);
     void ScheduleCopies(MachineBasicBlock* MBB, std::set<unsigned>& pushed);
     void InsertCopies(MachineBasicBlock* MBB,
                       SmallPtrSet<MachineBasicBlock*, 16>& v);
-    void mergeLiveIntervals(unsigned primary, unsigned secondary, unsigned VN);
+    void mergeLiveIntervals(unsigned primary, unsigned secondary, 
+                            MachineBasicBlock* pred);
   };
-
-  char StrongPHIElimination::ID = 0;
-  RegisterPass<StrongPHIElimination> X("strong-phi-node-elimination",
-                  "Eliminate PHI nodes for register allocation, intelligently");
 }
 
-const PassInfo *llvm::StrongPHIEliminationID = X.getPassInfo();
+char StrongPHIElimination::ID = 0;
+static RegisterPass<StrongPHIElimination>
+X("strong-phi-node-elimination",
+  "Eliminate PHI nodes for register allocation, intelligently");
+
+const PassInfo *const llvm::StrongPHIEliminationID = &X;
 
 /// computeDFS - Computes the DFS-in and DFS-out numbers of the dominator tree
 /// of the given MachineFunction.  These numbers are then used in other parts
@@ -192,6 +197,8 @@ void StrongPHIElimination::computeDFS(MachineFunction& MF) {
   }
 }
 
+namespace {
+
 /// PreorderSorter - a helper class that is used to sort registers
 /// according to the preorder number of their defining blocks
 class PreorderSorter {
@@ -219,10 +226,13 @@ public:
   }
 };
 
+}
+
 /// computeDomForest - compute the subforest of the DomTree corresponding
 /// to the defining blocks of the registers in question
 std::vector<StrongPHIElimination::DomForestNode*>
-StrongPHIElimination::computeDomForest(std::map<unsigned, unsigned>& regs, 
+StrongPHIElimination::computeDomForest(
+                  std::map<unsigned, MachineBasicBlock*>& regs, 
                                        MachineRegisterInfo& MRI) {
   // Begin by creating a virtual root node, since the actual results
   // may well be a forest.  Assume this node has maximum DFS-out number.
@@ -232,8 +242,8 @@ StrongPHIElimination::computeDomForest(std::map<unsigned, unsigned>& regs,
   // Populate a worklist with the registers
   std::vector<unsigned> worklist;
   worklist.reserve(regs.size());
-  for (std::map<unsigned, unsigned>::iterator I = regs.begin(), E = regs.end();
-       I != E; ++I)
+  for (std::map<unsigned, MachineBasicBlock*>::iterator I = regs.begin(),
+       E = regs.end(); I != E; ++I)
     worklist.push_back(I->first);
   
   // Sort the registers by the DFS-in number of their defining block
@@ -408,7 +418,7 @@ void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
   
   // Iterate over all the PHI nodes in this block
   MachineBasicBlock::iterator P = MBB->begin();
-  while (P->getOpcode() == TargetInstrInfo::PHI) {
+  while (P != MBB->end() && P->getOpcode() == TargetInstrInfo::PHI) {
     unsigned DestReg = P->getOperand(0).getReg();
 
     // Don't both doing PHI elimination for dead PHI's.
@@ -426,7 +436,7 @@ void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
     // are going to be renames rather than having copies inserted.  This set
     // is refinded over the course of this function.  UnionedBlocks is the set
     // of corresponding MBBs.
-    std::map<unsigned, unsigned> PHIUnion;
+    std::map<unsigned, MachineBasicBlock*> PHIUnion;
     SmallPtrSet<MachineBasicBlock*, 8> UnionedBlocks;
   
     // Iterate over the operands of the PHI node
@@ -460,13 +470,7 @@ void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
         UsedByAnother.insert(SrcReg);
       } else {
         // Otherwise, add it to the renaming set
-        LiveInterval& I = LI.getOrCreateInterval(SrcReg);
-        unsigned idx = LI.getMBBEndIdx(P->getOperand(i).getMBB());
-        VNInfo* VN = I.getLiveRangeContaining(idx)->valno;
-        
-        assert(VN && "No VNInfo for register?");
-        
-        PHIUnion.insert(std::make_pair(SrcReg, VN->id));
+        PHIUnion.insert(std::make_pair(SrcReg,P->getOperand(i).getMBB()));
         UnionedBlocks.insert(MRI.getVRegDef(SrcReg)->getParent());
       }
     }
@@ -483,8 +487,17 @@ void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
     std::vector<std::pair<unsigned, unsigned> > localInterferences;
     processPHIUnion(P, PHIUnion, DF, localInterferences);
     
+    // If one of the inputs is defined in the same block as the current PHI
+    // then we need to check for a local interference between that input and
+    // the PHI.
+    for (std::map<unsigned, MachineBasicBlock*>::iterator I = PHIUnion.begin(),
+         E = PHIUnion.end(); I != E; ++I)
+      if (MRI.getVRegDef(I->first)->getParent() == P->getParent())
+        localInterferences.push_back(std::make_pair(I->first,
+                                                    P->getOperand(0).getReg()));
+    
     // The dominator forest walk may have returned some register pairs whose
-    // interference cannot be determines from dominator analysis.  We now 
+    // interference cannot be determined from dominator analysis.  We now 
     // examine these pairs for local interferences.
     for (std::vector<std::pair<unsigned, unsigned> >::iterator I =
         localInterferences.begin(), E = localInterferences.end(); I != E; ++I) {
@@ -527,12 +540,12 @@ void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
       }
     }
     
-    // Add the renaming set for this PHI node to our overal renaming information
+    // Add the renaming set for this PHI node to our overall renaming information
     RenameSets.insert(std::make_pair(P->getOperand(0).getReg(), PHIUnion));
     
     // Remember which registers are already renamed, so that we don't try to 
     // rename them for another PHI node in this block
-    for (std::map<unsigned, unsigned>::iterator I = PHIUnion.begin(),
+    for (std::map<unsigned, MachineBasicBlock*>::iterator I = PHIUnion.begin(),
          E = PHIUnion.end(); I != E; ++I)
       ProcessedNames.insert(I->first);
     
@@ -545,7 +558,7 @@ void StrongPHIElimination::processBlock(MachineBasicBlock* MBB) {
 /// that are known to interfere, and flag others that need to be checked for
 /// local interferences.
 void StrongPHIElimination::processPHIUnion(MachineInstr* Inst,
-                                        std::map<unsigned, unsigned>& PHIUnion,
+                        std::map<unsigned, MachineBasicBlock*>& PHIUnion,
                         std::vector<StrongPHIElimination::DomForestNode*>& DF,
                         std::vector<std::pair<unsigned, unsigned> >& locals) {
   
@@ -617,7 +630,7 @@ void StrongPHIElimination::processPHIUnion(MachineInstr* Inst,
 /// of Static Single Assignment Form" by Briggs, et al.
 void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
                                           std::set<unsigned>& pushed) {
-  // FIXME: This function needs to update LiveVariables
+  // FIXME: This function needs to update LiveIntervals
   std::map<unsigned, unsigned>& copy_set= Waiting[MBB];
   
   std::map<unsigned, unsigned> worklist;
@@ -646,6 +659,8 @@ void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
   MachineRegisterInfo& MRI = MF->getRegInfo();
   const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
   
+  SmallVector<std::pair<unsigned, MachineInstr*>, 4> InsertedPHIDests;
+  
   // Iterate over the worklist, inserting copies
   while (!worklist.empty() || !copy_set.empty()) {
     while (!worklist.empty()) {
@@ -676,6 +691,11 @@ void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
                         map[curr.first], RC, RC);
       map[curr.first] = curr.second;
       
+      // Push this copy onto InsertedPHICopies so we can
+      // update LiveIntervals with it.
+      MachineBasicBlock::iterator MI = MBB->getFirstTerminator();
+      InsertedPHIDests.push_back(std::make_pair(curr.second, --MI));
+      
       // If curr.first is a destination in copy_set...
       for (std::map<unsigned, unsigned>::iterator I = copy_set.begin(),
            E = copy_set.end(); I != E; )
@@ -708,6 +728,21 @@ void StrongPHIElimination::ScheduleCopies(MachineBasicBlock* MBB,
       worklist.insert(curr);
     }
   }
+  
+  // Renumber the instructions so that we can perform the index computations
+  // needed to create new live intervals.
+  LI.computeNumbering();
+  
+  // For copies that we inserted at the ends of predecessors, we construct
+  // live intervals.  This is pretty easy, since we know that the destination
+  // register cannot have be in live at that point previously.  We just have
+  // to make sure that, for registers that serve as inputs to more than one
+  // PHI, we don't create multiple overlapping live intervals.
+  std::set<unsigned> RegHandled;
+  for (SmallVector<std::pair<unsigned, MachineInstr*>, 4>::iterator I =
+       InsertedPHIDests.begin(), E = InsertedPHIDests.end(); I != E; ++I)
+    if (RegHandled.insert(I->first).second)
+      LI.addLiveRangeToEndOfBlock(I->first, I->second);
 }
 
 /// InsertCopies - insert copies into MBB and all of its successors
@@ -742,110 +777,33 @@ void StrongPHIElimination::InsertCopies(MachineBasicBlock* MBB,
     Stacks[*I].pop_back();
 }
 
-/// ComputeUltimateVN - Assuming we are going to join two live intervals,
-/// compute what the resultant value numbers for each value in the input two
-/// ranges will be.  This is complicated by copies between the two which can
-/// and will commonly cause multiple value numbers to be merged into one.
-///
-/// VN is the value number that we're trying to resolve.  InstDefiningValue
-/// keeps track of the new InstDefiningValue assignment for the result
-/// LiveInterval.  ThisFromOther/OtherFromThis are sets that keep track of
-/// whether a value in this or other is a copy from the opposite set.
-/// ThisValNoAssignments/OtherValNoAssignments keep track of value #'s that have
-/// already been assigned.
-///
-/// ThisFromOther[x] - If x is defined as a copy from the other interval, this
-/// contains the value number the copy is from.
-///
-static unsigned ComputeUltimateVN(VNInfo *VNI,
-                                  SmallVector<VNInfo*, 16> &NewVNInfo,
-                                  DenseMap<VNInfo*, VNInfo*> &ThisFromOther,
-                                  DenseMap<VNInfo*, VNInfo*> &OtherFromThis,
-                                  SmallVector<int, 16> &ThisValNoAssignments,
-                                  SmallVector<int, 16> &OtherValNoAssignments) {
-  unsigned VN = VNI->id;
-
-  // If the VN has already been computed, just return it.
-  if (ThisValNoAssignments[VN] >= 0)
-    return ThisValNoAssignments[VN];
-//  assert(ThisValNoAssignments[VN] != -2 && "Cyclic case?");
-
-  // If this val is not a copy from the other val, then it must be a new value
-  // number in the destination.
-  DenseMap<VNInfo*, VNInfo*>::iterator I = ThisFromOther.find(VNI);
-  if (I == ThisFromOther.end()) {
-    NewVNInfo.push_back(VNI);
-    return ThisValNoAssignments[VN] = NewVNInfo.size()-1;
-  }
-  VNInfo *OtherValNo = I->second;
-
-  // Otherwise, this *is* a copy from the RHS.  If the other side has already
-  // been computed, return it.
-  if (OtherValNoAssignments[OtherValNo->id] >= 0)
-    return ThisValNoAssignments[VN] = OtherValNoAssignments[OtherValNo->id];
-  
-  // Mark this value number as currently being computed, then ask what the
-  // ultimate value # of the other value is.
-  ThisValNoAssignments[VN] = -2;
-  unsigned UltimateVN =
-    ComputeUltimateVN(OtherValNo, NewVNInfo, OtherFromThis, ThisFromOther,
-                      OtherValNoAssignments, ThisValNoAssignments);
-  return ThisValNoAssignments[VN] = UltimateVN;
-}
-
 void StrongPHIElimination::mergeLiveIntervals(unsigned primary,
                                               unsigned secondary,
-                                              unsigned secondaryVN) {
+                                              MachineBasicBlock* pred) {
   
   LiveIntervals& LI = getAnalysis<LiveIntervals>();
   LiveInterval& LHS = LI.getOrCreateInterval(primary);
   LiveInterval& RHS = LI.getOrCreateInterval(secondary);
   
-  // Compute the final value assignment, assuming that the live ranges can be
-  // coalesced.
-  SmallVector<int, 16> LHSValNoAssignments;
-  SmallVector<int, 16> RHSValNoAssignments;
-  SmallVector<VNInfo*, 16> NewVNInfo;
+  LI.computeNumbering();
   
-  LHSValNoAssignments.resize(LHS.getNumValNums(), -1);
-  RHSValNoAssignments.resize(RHS.getNumValNums(), -1);
-  NewVNInfo.reserve(LHS.getNumValNums() + RHS.getNumValNums());
+  const LiveRange* RangeMergingIn =
+                   RHS.getLiveRangeContaining(LI.getMBBEndIdx(pred));
+  VNInfo* NewVN = LHS.getNextValue(RangeMergingIn->valno->def,
+                                   RangeMergingIn->valno->copy,
+                                   LI.getVNInfoAllocator());
+  NewVN->hasPHIKill = true;
+  LiveRange NewRange(RangeMergingIn->start, RangeMergingIn->end, NewVN);
   
-  for (LiveInterval::vni_iterator I = LHS.vni_begin(), E = LHS.vni_end();
-       I != E; ++I) {
-    VNInfo *VNI = *I;
-    unsigned VN = VNI->id;
-    if (LHSValNoAssignments[VN] >= 0 || VNI->def == ~1U) 
-      continue;
-    
-    NewVNInfo.push_back(VNI);
-    LHSValNoAssignments[VN] = NewVNInfo.size()-1;
-  }
+  if (RHS.containsOneValue())
+    LI.removeInterval(RHS.reg);
+  else
+    RHS.removeRange(RangeMergingIn->start, RangeMergingIn->end, true);
   
-  for (LiveInterval::vni_iterator I = RHS.vni_begin(), E = RHS.vni_end();
-       I != E; ++I) {
-    VNInfo *VNI = *I;
-    unsigned VN = VNI->id;
-    if (RHSValNoAssignments[VN] >= 0 || VNI->def == ~1U)
-      continue;
-      
-    NewVNInfo.push_back(VNI);
-    RHSValNoAssignments[VN] = NewVNInfo.size()-1;
-  }
-
-  // If we get here, we know that we can coalesce the live ranges.  Ask the
-  // intervals to coalesce themselves now.
-
-  LHS.join(RHS, &LHSValNoAssignments[0], &RHSValNoAssignments[0], NewVNInfo);
-  LI.removeInterval(secondary);
-  
-  // The valno that was previously the input to the PHI node
-  // now has a PHIKill.
-  LHS.getValNumInfo(RHSValNoAssignments[secondaryVN])->hasPHIKill = true;
+  LHS.addRange(NewRange);
 }
 
 bool StrongPHIElimination::runOnMachineFunction(MachineFunction &Fn) {
-  
   LiveIntervals& LI = getAnalysis<LiveIntervals>();
   
   // Compute DFS numbers of each block
@@ -858,16 +816,17 @@ bool StrongPHIElimination::runOnMachineFunction(MachineFunction &Fn) {
       processBlock(I);
   
   // Insert copies
-  // FIXME: This process should probably preserve LiveVariables
+  // FIXME: This process should probably preserve LiveIntervals
   SmallPtrSet<MachineBasicBlock*, 16> visited;
   InsertCopies(Fn.begin(), visited);
   
   // Perform renaming
-  typedef std::map<unsigned, std::map<unsigned, unsigned> > RenameSetType;
+  typedef std::map<unsigned, std::map<unsigned, MachineBasicBlock*> >
+          RenameSetType;
   for (RenameSetType::iterator I = RenameSets.begin(), E = RenameSets.end();
        I != E; ++I)
-    for (std::map<unsigned, unsigned>::iterator SI = I->second.begin(),
-         SE = I->second.end(); SI != SE; ++SI) {
+    for (std::map<unsigned, MachineBasicBlock*>::iterator SI = 
+         I->second.begin(), SE = I->second.end(); SI != SE; ++SI) {
       mergeLiveIntervals(I->first, SI->first, SI->second);
       Fn.getRegInfo().replaceRegWith(SI->first, I->first);
     }
@@ -887,22 +846,45 @@ bool StrongPHIElimination::runOnMachineFunction(MachineFunction &Fn) {
        I != E; ) {
     MachineInstr* PInstr = *(I++);
     
+    // Trim live intervals of input registers.  They are no longer live into
+    // this block.
+    for (unsigned i = 1; i < PInstr->getNumOperands(); i += 2) {
+      unsigned reg = PInstr->getOperand(i).getReg();
+      MachineBasicBlock* MBB = PInstr->getOperand(i+1).getMBB();
+      LiveInterval& InputI = LI.getInterval(reg);
+      if (MBB != PInstr->getParent() &&
+          InputI.liveAt(LI.getMBBStartIdx(PInstr->getParent())))
+        InputI.removeRange(LI.getMBBStartIdx(PInstr->getParent()),
+                           LI.getInstructionIndex(PInstr),
+                           true);
+    }
+    
     // If this is a dead PHI node, then remove it from LiveIntervals.
     unsigned DestReg = PInstr->getOperand(0).getReg();
+    LiveInterval& PI = LI.getInterval(DestReg);
     if (PInstr->registerDefIsDead(DestReg)) {
-      LiveInterval& PI = LI.getInterval(DestReg);
-      
       if (PI.containsOneValue()) {
         LI.removeInterval(DestReg);
       } else {
         unsigned idx = LI.getDefIndex(LI.getInstructionIndex(PInstr));
         PI.removeRange(*PI.getLiveRangeContaining(idx), true);
       }
+    } else {
+      // If the PHI is not dead, then the valno defined by the PHI
+      // now has an unknown def.
+      unsigned idx = LI.getDefIndex(LI.getInstructionIndex(PInstr));
+      const LiveRange* PLR = PI.getLiveRangeContaining(idx);
+      PLR->valno->def = ~0U;
+      LiveRange R (LI.getMBBStartIdx(PInstr->getParent()),
+                   PLR->start, PLR->valno);
+      PI.addRange(R);
     }
-      
+    
     LI.RemoveMachineInstrFromMaps(PInstr);
     PInstr->eraseFromParent();
   }
   
+  LI.computeNumbering();
+  
   return true;
 }