Create type IDs based on reference counts.

lib/Bitcode/NaCl/Writer/NaClValueEnumerator.h

Index: lib/Bitcode/NaCl/Writer/NaClValueEnumerator.h
diff --git a/lib/Bitcode/NaCl/Writer/NaClValueEnumerator.h b/lib/Bitcode/NaCl/Writer/NaClValueEnumerator.h
index 9e9954883a61406ab6a946e552bf6ab93aed6a56..7fcfd27199e1a2b15a536838de21991ca964d8aa 100644
--- a/lib/Bitcode/NaCl/Writer/NaClValueEnumerator.h
+++ b/lib/Bitcode/NaCl/Writer/NaClValueEnumerator.h
@@ -42,8 +42,17 @@ public:
   // For each value, we remember its Value* and occurrence frequency.
   typedef std::vector<std::pair<const Value*, unsigned> > ValueList;
 private:
+  // Defines unique ID's for each type.
   typedef DenseMap<Type*, unsigned> TypeMapType;
   TypeMapType TypeMap;
+  // Defines the number of references to each type. If defined,
+  // we are in the first pass of collecting types, and reference counts
+  // should be added to the map. If undefined, we are in the second pass
+  // that actually assigns type IDs, based on frequency counts found in
+  // the first pass.
+  typedef TypeMapType TypeCountMapType;
+  TypeCountMapType* TypeCountMap;

[jvoung (off chromium), 2013/04/29 18:35:42]

Does this need to be a pointer, or could it be allocated as part of the
ValueEnumerator, like TypeMap, ValueMap, etc.?

[Karl, 2013/04/29 22:19:16]

I did it this way because it is only needed during the first pass of the
construction of the type ID map. Once built, this space is wasted and should be
collected.

+
   TypeList Types;
 
   typedef DenseMap<const Value*, unsigned> ValueMapType;
@@ -83,6 +92,7 @@ private:
 
   unsigned FirstFuncConstantID;
   unsigned FirstInstID;
+  unsigned FirstFloatTypeID;

[jvoung (off chromium), 2013/04/29 18:35:42]

why are float types singled out?

[Karl, 2013/04/29 22:19:16]

I was looking ahead to "float" operations, which can subtract this value from
the type id. However, since I'm not adding that, I will remove it until it is
needed.

 
   NaClValueEnumerator(const NaClValueEnumerator &) LLVM_DELETED_FUNCTION;
   void operator=(const NaClValueEnumerator &) LLVM_DELETED_FUNCTION;
@@ -152,6 +162,7 @@ public:
   void purgeFunction();
 
 private:
+  void OptimizeTypes(const Module *M);

[jvoung (off chromium), 2013/04/29 18:35:42]

This seems like a change that could be upstreamed.  It's not actually changing
the wireformat, just optimizing the choice of IDs?

I'll review this, but just saying that if we upstream, it could reduce the
number of differences.

[Karl, 2013/04/29 22:19:16]

I agree that it is worth considering upstreaming this change.

   void OptimizeConstants(unsigned CstStart, unsigned CstEnd);
 
   void EnumerateMDNodeOperands(const MDNode *N);