[RewriteStatepointsForGC] Make base pointer inference deterministic

[reames]

Summary

Previously, the base pointer algorithm wasn't deterministic. The core fixed point was (of course), but we were inserting new nodes and optimizing them in an order which was unspecified and variable. We'd somewhat hacked around this for testing by sorting by value name, but that doesn't solve the general determinism problem.

Instead, we can use the order of traversal over the def/use graph to give us a single consistent ordering. Today, this is a DFS order, but the exact order doesn't mater provided it's deterministic for a given input.

(Q: It is safe to rely on a deterministic order of operands right?)

Note that this only fixes the determinism within a single inference step. The inference step is currently invoked many times in a non-deterministic order. That's the next change in the sequence. :)

[sanjoy]

I'm not a 100% sure about this, but I think we can just use a MapVector for the states variable and then not track the extra VisitOrder. If you've already considered using MapVector and it does not work for some reason then this change lgtm, otherwise please use MapVector.

lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

817

Can we just typedef MapVector<Value *, BDVState> ConflictStateMapTy; instead, and then instead of maintaining an extra VisitOrder insert into states instead of VisitOrder.insert(Base)?

986

The naming here is weird -- how about Instruction *I = ...? Fine to do as a follow up change.

987

In that follow up change, I think state should be renamed to State in keeping with LLVM coding style.

[reames]

Update per Sanjoy's comments. Naming changes will follow in separate change.

[sanjoy]

lgtm

lib/Transforms/Scalar/RewriteStatepointsForGC.cpp

static Value *findBasePointer(Value *I, DefiningValueMapTy &cache) {
   // overall worse solution.
 
 #ifndef NDEBUG
   auto isExpectedBDVType = [](Value *BDV) {
     return isa<PHINode>(BDV) || isa<SelectInst>(BDV) || isa<ExtractElementInst>(BDV);
   };
 #endif
 
-  // Once populated, will contain a mapping from each potentially non-base BDV
-  // to a lattice value (described above) which corresponds to that BDV.
-  ConflictStateMapTy states;
+  // Once populated, will contain the list of base defining values we need to
+  // analyze and possibly insert parallel base propagation instructions for.
+  // We use the order of insertion (DFS over the def/use graph) to provide a
+  // stable deterministic ordering for visiting DenseMaps (which are unordered)
+  // below.  This is important for deterministic compilation.
+  SmallSetVector<Value *, 16> VisitOrder;
+  
   // Recursively fill in all phis & selects reachable from the initial one
   // for which we don't already know a definite base value for
   /* scope */ {
-    DenseSet<Value *> Visited;
     SmallVector<Value*, 16> Worklist;
     Worklist.push_back(def);
-    Visited.insert(def);
+    VisitOrder.insert(def);
     while (!Worklist.empty()) {
       Value *Current = Worklist.pop_back_val();
       assert(!isKnownBaseResult(Current) && "why did it get added?");
 
       auto visitIncomingValue = [&](Value *InVal) {
         Value *Base = findBaseOrBDV(InVal, cache);
         if (isKnownBaseResult(Base))
           // Known bases won't need new instructions introduced and can be
           // ignored safely
           return;
         assert(isExpectedBDVType(Base) && "the only non-base values "
                "we see should be base defining values");
-        if (Visited.insert(Base).second)
+        if (VisitOrder.insert(Base))
           Worklist.push_back(Base);
       };
       if (PHINode *Phi = dyn_cast<PHINode>(Current)) {
         for (Value *InVal : Phi->incoming_values())
           visitIncomingValue(InVal);
       } else if (SelectInst *Sel = dyn_cast<SelectInst>(Current)) {
         visitIncomingValue(Sel->getTrueValue());
         visitIncomingValue(Sel->getFalseValue());
       } else if (auto *EE = dyn_cast<ExtractElementInst>(Current)) {
         visitIncomingValue(EE->getVectorOperand());
       } else {
         // There are two classes of instructions we know we don't handle.
         assert(isa<ShuffleVectorInst>(Current) ||
                isa<InsertElementInst>(Current));
         llvm_unreachable("unimplemented instruction case");
       }
     }
-    // The frontier of visited instructions are the ones we might need to
-    // duplicate, so fill in the starting state for the optimistic algorithm
-    // that follows.
-    for (Value *BDV : Visited) {
-      states[BDV] = BDVState();
-    }
-  }
+  }
+
+  // Once populated, will contain a mapping from each potentially non-base BDV
+  // to a lattice value (described above) which corresponds to that BDV.
+  ConflictStateMapTy states;
+
+  for (Value *BDV : VisitOrder) {
 
 #ifndef NDEBUG
   DEBUG(dbgs() << "States after initialization:\n");
-  for (auto Pair : states) {
-    DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
+  for (auto Key : VisitOrder) {
+    auto State = states[Key];
+    DEBUG(dbgs() << " " << State << " for " << *Key << "\n");
   }
 #endif
 
   // Return a phi state for a base defining value.  We'll generate a new
   // base state for known bases and expect to find a cached state otherwise.
   auto getStateForBDV = [&](Value *baseValue) {
     if (isKnownBaseResult(baseValue))
       return BDVState(baseValue);
     auto I = states.find(baseValue);
     assert(I != states.end() && "lookup failed!");
     return I->second;
   };
 
   bool progress = true;
   while (progress) {
 #ifndef NDEBUG
     size_t oldSize = states.size();
 #endif
     progress = false;
-    // We're only changing keys in this loop, thus safe to keep iterators
+    // We're only changing values in this loop, thus safe to keep iterators.
+    // Since this is computing a fixed point, the order of visit does not
+    // effect the result.  TODO: We could use a worklist here and make this run
+    // much faster.
     for (auto Pair : states) {
       Value *v = Pair.first;
       assert(!isKnownBaseResult(v) && "why did it get added?");
 
       // Given an input value for the current instruction, return a BDVState
       // instance which represents the BDV of that value.
       auto getStateForInput = [&](Value *V) mutable {
         Value *BDV = findBaseOrBDV(V, cache);
for (auto Pair : states) {
           calculateMeet.meetWith(getStateForInput(Val));
       } else {
         // The 'meet' for an extractelement is slightly trivial, but it's still
         // useful in that it drives us to conflict if our input is.
         auto *EE = cast<ExtractElementInst>(v);
         calculateMeet.meetWith(getStateForInput(EE->getVectorOperand()));
       }
 
-
       BDVState oldState = states[v];
       BDVState newState = calculateMeet.getResult();
       if (oldState != newState) {
         progress = true;
         states[v] = newState;
       }
     }
 
     assert(oldSize <= states.size());
     assert(oldSize == states.size() || progress);
   }
 
 #ifndef NDEBUG
   DEBUG(dbgs() << "States after meet iteration:\n");
-  for (auto Pair : states) {
-    DEBUG(dbgs() << " " << Pair.second << " for " << *Pair.first << "\n");
+  for (auto Key : VisitOrder) {
+    auto State = states[Key];
+    DEBUG(dbgs() << " " << State << " for " << *Key << "\n");
   }
 #endif
   
   // Insert Phis for all conflicts
   // We want to keep naming deterministic in the loop that follows, so
   // sort the keys before iteration.  This is useful in allowing us to
   // write stable tests. Note that there is no invalidation issue here.
-  SmallVector<Value *, 16> Keys;
-  Keys.reserve(states.size());
-  for (auto Pair : states) {
-    Value *V = Pair.first;
-    Keys.push_back(V);
-  }
-  std::sort(Keys.begin(), Keys.end(), order_by_name);
   // TODO: adjust naming patterns to avoid this order of iteration dependency
-  for (Value *V : Keys) {
+  for (Value *V : VisitOrder) {
     Instruction *I = cast<Instruction>(V);
     BDVState State = states[I];
     assert(!isKnownBaseResult(I) && "why did it get added?");
     assert(!State.isUnknown() && "Optimistic algorithm didn't complete!");
 
     // extractelement instructions are a bit special in that we may need to
     // insert an extract even when we know an exact base for the instruction.
     // The problem is that we need to convert from a vector base to a scalar
auto getBaseForInput = [&](Value *Input, Instruction *InsertPt) {
     if (Base->getType() != Input->getType() &&
         InsertPt) {
       Base = new BitCastInst(Base, Input->getType(), "cast",
                              InsertPt);
     }
     return Base;
   };
 
-  // Fixup all the inputs of the new PHIs
-  for (auto Pair : states) {
-    Instruction *v = cast<Instruction>(Pair.first);
-    BDVState state = Pair.second;
+  // Fixup all the inputs of the new PHIs.  Visit order needs to be
+  // deterministic and predictable because we're naming newly created
+  // instructions.
+  for (auto *Key : VisitOrder) {
+    Instruction *v = cast<Instruction>(Key);
+    BDVState state = states[Key];
 
     assert(!isKnownBaseResult(v) && "why did it get added?");
     assert(!state.isUnknown() && "Optimistic algorithm didn't complete!");
     if (!state.isConflict())
       continue;
 
     if (PHINode *basephi = dyn_cast<PHINode>(state.getBase())) {
       PHINode *phi = cast<PHINode>(v);
#endif
   // doing as a post process step is easier to reason about for the moment.
   DenseMap<Value *, Value *> ReverseMap;
   SmallPtrSet<Instruction *, 16> NewInsts;
   SmallSetVector<AssertingVH<Instruction>, 16> Worklist;
   // Note: We need to visit the states in a deterministic order.  We uses the
   // Keys we sorted above for this purpose.  Note that we are papering over a
   // bigger problem with the algorithm above - it's visit order is not
   // deterministic.  A larger change is needed to fix this.
-  for (auto Key : Keys) {
-    Value *V = Key;
-    auto State = states[Key];
+  for (auto *BDV : VisitOrder) {
+    auto State = states[BDV];
     Value *Base = State.getBase();
-    assert(V && Base);
-    assert(!isKnownBaseResult(V) && "why did it get added?");
+    assert(BDV && Base);
+    assert(!isKnownBaseResult(BDV) && "why did it get added?");
     assert(isKnownBaseResult(Base) &&
            "must be something we 'know' is a base pointer");
     if (!State.isConflict())
       continue;
 
-    ReverseMap[Base] = V;
+    ReverseMap[Base] = BDV;
     if (auto *BaseI = dyn_cast<Instruction>(Base)) {
       NewInsts.insert(BaseI);
       Worklist.insert(BaseI);
     }
   }
   auto ReplaceBaseInstWith = [&](Value *BDV, Instruction *BaseI,
                                  Value *Replacement) {
     // Add users which are new instructions (excluding self references)
if (Value *V = SimplifyInstruction(BaseI, DL)) {
       ReplaceBaseInstWith(Bdv, BaseI, V);
       continue;
     }
   }
 
   // Cache all of our results so we can cheaply reuse them
   // NOTE: This is actually two caches: one of the base defining value
   // relation and one of the base pointer relation!  FIXME
-  for (auto item : states) {
-    Value *v = item.first;
-    Value *base = item.second.getBase();
-    assert(v && base);
+  for (auto *BDV : VisitOrder) {
+    Value *base = states[BDV].getBase();
+    assert(BDV && base);
 
     std::string fromstr =
-      cache.count(v) ? (cache[v]->hasName() ? cache[v]->getName() : "")
+      cache.count(BDV) ? (cache[BDV]->hasName() ? cache[BDV]->getName() : "")
                      : "none";
     DEBUG(dbgs() << "Updating base value cache"
-          << " for: " << (v->hasName() ? v->getName() : "")
+          << " for: " << (BDV->hasName() ? BDV->getName() : "")
           << " from: " << fromstr
           << " to: " << (base->hasName() ? base->getName() : "") << "\n");
 
-    if (cache.count(v)) {
+    if (cache.count(BDV)) {
       // Once we transition from the BDV relation being store in the cache to
       // the base relation being stored, it must be stable
-      assert((!isKnownBaseResult(cache[v]) || cache[v] == base) &&
+      assert((!isKnownBaseResult(cache[BDV]) || cache[BDV] == base) &&
              "base relation should be stable");
     }
-    cache[v] = base;
+    cache[BDV] = base;
   }
   assert(cache.find(def) != cache.end());
   return cache[def];
 }
 
 // For a set of live pointers (base and/or derived), identify the base
 // pointer of the object which they are derived from.  This routine will
 // mutate the IR graph as needed to make the 'base' pointer live at the