Fix for PR14824: "Optimization arm_ldst_opt inserts newly generated instruction vldmia at incorrect position".

Patch introduces memory operands tracking in ARMLoadStoreOpt::LoadStoreMultipleOpti. For each register it keeps the order of load operations as it was before optimization pass.
It is kind of deep improvement of fix proposed by Hao: http://llvm.org/bugs/show_bug.cgi?id=14824#c4
But it also tracks conflicts between different register classes (e.g. D2 and S5).
For more details see:
Bug description: http://llvm.org/bugs/show_bug.cgi?id=14824
LLVM Commits discussion: 
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20130311/167936.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20130318/168688.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20130325/169376.html
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20130401/170238.html



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178851 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 161 insertions, 10 deletions

diff --git a/lib/Target/ARM/ARMLoadStoreOptimizer.cpp b/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
index e4e683c2a0..0682459e1c 100644
--- a/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
+++ b/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
@@ -87,6 +87,53 @@ namespace {
                       MachineBasicBlock::iterator i)
         : Offset(o), Reg(r), isKill(k), Position(p), MBBI(i), Merged(false) {}
     };
+    class UnitRegsMap {
+    public:
+      UnitRegsMap(const TargetRegisterInfo* _TRI) : TRI(_TRI) {}
+      const SmallVector<unsigned, 4>& operator[](unsigned Reg) {
+        DenseMap<unsigned, SmallVector<unsigned, 4> >::iterator found =
+            Cache.find(Reg);
+        if (found != Cache.end())
+          return found->second;
+        else
+          return Cache.insert(std::make_pair(Reg, this->getUnitRegs(Reg)))
+                      .first->second;
+      }
+    private:
+      SmallVector<unsigned, 4> getUnitRegs(unsigned Reg) {
+        SmallVector<unsigned, 4> Res;
+
+        const TargetRegisterClass* TRC = TRI->getRegClass(Reg);
+        if (TRC == &ARM::QPRRegClass) {
+          if (Reg > ARM::Q7) {
+            Res.push_back(TRI->getSubReg(Reg, ARM::dsub_0));
+            Res.push_back(TRI->getSubReg(Reg, ARM::dsub_1));
+            return Res;
+          }
+
+          Res.push_back(TRI->getSubReg(Reg, ARM::ssub_0));
+          Res.push_back(TRI->getSubReg(Reg, ARM::ssub_1));
+          Res.push_back(TRI->getSubReg(Reg, ARM::ssub_2));
+          Res.push_back(TRI->getSubReg(Reg, ARM::ssub_3));
+
+          return Res;
+        }
+
+        if (TRC == &ARM::DPRRegClass && Reg < ARM::D15) {
+          Res.push_back(TRI->getSubReg(Reg, ARM::ssub_0));
+          Res.push_back(TRI->getSubReg(Reg, ARM::ssub_1));
+
+          return Res;
+        }
+
+        Res.push_back(Reg);
+
+        return Res;
+
+      }
+      const TargetRegisterInfo* TRI;
+      DenseMap<unsigned, SmallVector<unsigned, 4> > Cache;
+    };
     typedef SmallVector<MemOpQueueEntry,8> MemOpQueue;
     typedef MemOpQueue::iterator MemOpQueueIter;
 
@@ -128,6 +175,11 @@ namespace {
                                    MachineBasicBlock::iterator MBBI,
                                    bool &Advance,
                                    MachineBasicBlock::iterator &I);
+    unsigned AddMemOp(MemOpQueue& MemOps,
+                      const MemOpQueueEntry newEntry,
+                      UnitRegsMap& UnitRegsInfo,
+                      SmallSet<unsigned, 4>& UsedUnitRegs,
+                      unsigned At = -1U);
     bool LoadStoreMultipleOpti(MachineBasicBlock &MBB);
     bool MergeReturnIntoLDM(MachineBasicBlock &MBB);
   };
@@ -1213,12 +1265,103 @@ bool ARMLoadStoreOpt::FixInvalidRegPairOp(MachineBasicBlock &MBB,
   return false;
 }
 
+/// AddMemOp - helper for ARMLoadStoreOpt::LoadStoreMultipleOpti.
+/// It adds store mem ops with simple push_back/insert method,
+/// without any additional logic.
+/// For load operation it does the next:
+/// 1. Adds new load operation into MemOp collection at "At" position.
+/// 2. Removes any "load" operations from MemOps, that changes "Reg" register
+/// contents, prior to "At".
+/// UnitRegsInfo - Map of type Map< Register, UnitRegisters-vector >
+/// UsedUnitRegs - set of unit-registers currently in use.
+/// At - position at which it would added, and prior which the clean-up
+/// should be made (for load operation).
+/// FIXME: The clean-up also should be made for store operations,
+/// but the memory address should be analyzed instead of unit registers.
+unsigned ARMLoadStoreOpt::AddMemOp(MemOpQueue& MemOps,
+                                   const MemOpQueueEntry NewEntry,
+                                   UnitRegsMap& UnitRegsInfo,
+                                   SmallSet<unsigned, 4>& UsedUnitRegs,
+                                   unsigned At) {
+  unsigned Cleaned = 0;
+
+  if (At == -1U) {
+    At = MemOps.size();
+    MemOps.push_back(NewEntry);
+  } else
+    MemOps.insert(&MemOps[At], NewEntry);
+
+  // FIXME:
+  // If operation is not load, leave it as is by now,
+  // So 0 overridden ops would cleaned in this case.
+  if (!NewEntry.MBBI->mayLoad())
+    return 0;
+
+  const SmallVector<unsigned, 4>& NewEntryUnitRegs = UnitRegsInfo[NewEntry.Reg];
+
+  bool FoundOverriddenLoads = false;
+
+  for (unsigned i = 0, e = NewEntryUnitRegs.size(); i != e; ++i)
+    if (UsedUnitRegs.count(NewEntryUnitRegs[i])) {
+      FoundOverriddenLoads = true;
+      break;
+    }
+
+  // If we detect that this register is used by load operations that are
+  // predecessors for the new one, remove them from MemOps then.
+  if (FoundOverriddenLoads) {
+    MemOpQueue UpdatedMemOps;
+
+    // Scan through MemOps entries.
+    for (unsigned i = 0; i != At; ++i) {
+      MemOpQueueEntry& MemOpEntry = MemOps[i];
+
+      // FIXME: Skip non-load operations by now.
+      if (!MemOpEntry.MBBI->mayLoad())
+        continue;
+
+      const SmallVector<unsigned, 4>& MemOpUnitRegs =
+          UnitRegsInfo[MemOpEntry.Reg];
+
+      // Lookup entry that loads contents into register used by new entry.
+      bool ReleaseThisEntry = false;
+      for (unsigned m = 0, em = MemOpUnitRegs.size(); m != em; ++m) {
+        if (std::find(NewEntryUnitRegs.begin(), NewEntryUnitRegs.end(),
+                      MemOpUnitRegs[m]) != NewEntryUnitRegs.end()) {
+          ReleaseThisEntry = true;
+          ++Cleaned;
+          break;
+        }
+      }
+
+      if (ReleaseThisEntry) {
+        const SmallVector<unsigned, 4>& RelesedRegs = UnitRegsInfo[MemOpEntry.Reg];
+        for (unsigned r = 0, er = RelesedRegs.size(); r != er; ++r)
+          UsedUnitRegs.erase(RelesedRegs[r]);
+      } else
+        UpdatedMemOps.push_back(MemOpEntry);
+    }
+
+    // Keep anything without changes after At position.
+    for (unsigned i = At, e = MemOps.size(); i != e; ++i)
+      UpdatedMemOps.push_back(MemOps[i]);
+
+    MemOps.swap(UpdatedMemOps);
+  }
+
+  UsedUnitRegs.insert(NewEntryUnitRegs.begin(), NewEntryUnitRegs.end());
+
+  return Cleaned;
+}
+
 /// LoadStoreMultipleOpti - An optimization pass to turn multiple LDR / STR
 /// ops of the same base and incrementing offset into LDM / STM ops.
 bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
   unsigned NumMerges = 0;
   unsigned NumMemOps = 0;
   MemOpQueue MemOps;
+  UnitRegsMap UnitRegsInfo(TRI);
+  SmallSet<unsigned, 4> UsedRegUnits;
   unsigned CurrBase = 0;
   int CurrOpc = -1;
   unsigned CurrSize = 0;
@@ -1265,8 +1408,11 @@ bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
         CurrSize = Size;
         CurrPred = Pred;
         CurrPredReg = PredReg;
+
         MemOps.push_back(MemOpQueueEntry(Offset, Reg, isKill, Position, MBBI));
         ++NumMemOps;
+        const SmallVector<unsigned, 4>& EntryUnitRegs = UnitRegsInfo[Reg];
+        UsedRegUnits.insert(EntryUnitRegs.begin(), EntryUnitRegs.end());
         Advance = true;
       } else {
         if (Clobber) {
@@ -1278,20 +1424,24 @@ bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
           // No need to match PredReg.
           // Continue adding to the queue.
           if (Offset > MemOps.back().Offset) {
-            MemOps.push_back(MemOpQueueEntry(Offset, Reg, isKill,
-                                             Position, MBBI));
-            ++NumMemOps;
+            unsigned OverridesCleaned =
+              AddMemOp(MemOps,
+                           MemOpQueueEntry(Offset, Reg, isKill, Position, MBBI),
+                           UnitRegsInfo, UsedRegUnits) != 0;
+            NumMemOps += 1 - OverridesCleaned;
             Advance = true;
           } else {
-            for (MemOpQueueIter I = MemOps.begin(), E = MemOps.end();
-                 I != E; ++I) {
-              if (Offset < I->Offset) {
-                MemOps.insert(I, MemOpQueueEntry(Offset, Reg, isKill,
-                                                 Position, MBBI));
-                ++NumMemOps;
+            for (unsigned I = 0; I != NumMemOps; ++I) {
+              if (Offset < MemOps[I].Offset) {
+                MemOpQueueEntry entry(Offset, Reg, isKill, Position, MBBI);
+                unsigned OverridesCleaned =
+                    AddMemOp(MemOps, entry, UnitRegsInfo,
+                                 UsedRegUnits, I) != 0;
+                NumMemOps += 1 - OverridesCleaned;
+
                 Advance = true;
                 break;
-              } else if (Offset == I->Offset) {
+              } else if (Offset == MemOps[I].Offset) {
                 // Collision! This can't be merged!
                 break;
               }
@@ -1362,6 +1512,7 @@ bool ARMLoadStoreOpt::LoadStoreMultipleOpti(MachineBasicBlock &MBB) {
       CurrPredReg = 0;
       if (NumMemOps) {
         MemOps.clear();
+        UsedRegUnits.clear();
         NumMemOps = 0;
       }