Changes to the V2 bytecode format:

- Null values are implicitly encoded instead of explicitly, this makes
    things more compact!
  - More compactly represent ConstantPointerRefs
  - Bytecode files are represented as:
      Header|GlobalTypes|GlobalVars/Function Protos|Constants|Functions|SymTab
    instead of
      Header|GlobalTypes|Constants|GlobalVars/Function Protos|Functions|SymTab
    which makes a lot of things simpler.

Writer changes:
  - We now explictly encode versioning information in the bytecode files.
  - This allows new code to read bytecode files produced by old code, but
    new bytecode files can have enhancements such as the above.  Although this
    makes the reader a bit more complex (having to deal with old formats), the
    writer only needs to be able to produce the most recent version.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@5749 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Bytecode/Writer/ConstantWriter.cpp

@@ -13,8 +13,6 @@
 #include "llvm/Constants.h"
 #include "llvm/SymbolTable.h"
 #include "llvm/DerivedTypes.h"
-#include <iostream>
-using std::cerr;
 
 void BytecodeWriter::outputType(const Type *T) {
   output_vbr((unsigned)T->getPrimitiveID(), Out);
@@ -52,7 +50,7 @@ void BytecodeWriter::outputType(const Type *T) {
     int Slot = Table.getValSlot(AT->getElementType());
     assert(Slot != -1 && "Type used but not available!!");
     output_vbr((unsigned)Slot, Out);
-    //cerr << "Type slot = " << Slot << " Type = " << T->getName() << endl;
+    //std::cerr << "Type slot = " << Slot << " Type = " << T->getName() << endl;
 
     output_vbr(AT->getNumElements(), Out);
     break;
@@ -89,13 +87,15 @@ void BytecodeWriter::outputType(const Type *T) {
 
   //case Type::PackedTyID:
   default:
-    cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
-	 << " Type '" << T->getDescription() << "'\n";
+    std::cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
+              << " Type '" << T->getDescription() << "'\n";
     break;
   }
 }
 
 bool BytecodeWriter::outputConstant(const Constant *CPV) {
+  assert((CPV->getType()->isPrimitiveType() || !CPV->isNullValue()) &&
+         "Shouldn't output null constants!");
 
   // We must check for a ConstantExpr before switching by type because
   // a ConstantExpr can be of any type, and has no explicit value.
@@ -121,9 +121,9 @@ bool BytecodeWriter::outputConstant(const Constant *CPV) {
   switch (CPV->getType()->getPrimitiveID()) {
   case Type::BoolTyID:    // Boolean Types
     if (cast<const ConstantBool>(CPV)->getValue())
-      output_vbr((unsigned)1, Out);
+      output_vbr(1U, Out);
     else
-      output_vbr((unsigned)0, Out);
+      output_vbr(0U, Out);
     break;
 
   case Type::UByteTyID:   // Unsigned integer types...
@@ -171,17 +171,11 @@ bool BytecodeWriter::outputConstant(const Constant *CPV) {
 
   case Type::PointerTyID: {
     const ConstantPointer *CPP = cast<const ConstantPointer>(CPV);
-    if (isa<ConstantPointerNull>(CPP)) {
-      output_vbr((unsigned)0, Out);
-    } else if (const ConstantPointerRef *CPR = 
-	                dyn_cast<ConstantPointerRef>(CPP)) {
-      output_vbr((unsigned)1, Out);
-      int Slot = Table.getValSlot((Value*)CPR->getValue());
-      assert(Slot != -1 && "Global used but not available!!");
-      output_vbr((unsigned)Slot, Out);
-    } else {
-      assert(0 && "Unknown ConstantPointer Subclass!");
-    }
+    assert(!isa<ConstantPointerNull>(CPP) && "Null should be already emitted!");
+    const ConstantPointerRef *CPR = cast<ConstantPointerRef>(CPP);
+    int Slot = Table.getValSlot((Value*)CPR->getValue());
+    assert(Slot != -1 && "Global used but not available!!");
+    output_vbr((unsigned)Slot, Out);
     break;
   }
 
@@ -199,8 +193,8 @@ bool BytecodeWriter::outputConstant(const Constant *CPV) {
   case Type::VoidTyID: 
   case Type::LabelTyID:
   default:
-    cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
-	 << " type '" << CPV->getType()->getName() << "'\n";
+    std::cerr << __FILE__ << ":" << __LINE__ << ": Don't know how to serialize"
+              << " type '" << CPV->getType()->getName() << "'\n";
     break;
   }
   return false;

lib/Bytecode/Writer/Writer.cpp

@@ -43,19 +43,36 @@ BytecodeWriter::BytecodeWriter(std::deque<unsigned char> &o, const Module *M)
   // Emit the top level CLASS block.
   BytecodeBlock ModuleBlock(BytecodeFormat::Module, Out);
 
-  // Output the ID of first "derived" type:
-  output_vbr((unsigned)Type::FirstDerivedTyID, Out);
+  bool isBigEndian = true;
+  bool hasLongPointers = true;
+
+  // Output the version identifier... we are currently on bytecode version #1
+  unsigned Version = (1 << 4) | isBigEndian | (hasLongPointers << 1);
+  output_vbr(Version, Out);
   align32(Out);
 
-  // Output module level constants, including types used by the function protos
-  outputConstants(false);
+  {
+    BytecodeBlock CPool(BytecodeFormat::GlobalTypePlane, Out);
+    
+    // Write the type plane for types first because earlier planes (e.g. for a
+    // primitive type like float) may have constants constructed using types
+    // coming later (e.g., via getelementptr from a pointer type).  The type
+    // plane is needed before types can be fwd or bkwd referenced.
+    const std::vector<const Value*> &Plane = Table.getPlane(Type::TypeTyID);
+    assert(!Plane.empty() && "No types at all?");
+    unsigned ValNo = Type::FirstDerivedTyID; // Start at the derived types...
+    outputConstantsInPlane(Plane, ValNo);      // Write out the types
+  }
 
-  // The ModuleInfoBlock follows directly after the Module constant pool
+  // The ModuleInfoBlock follows directly after the type information
   outputModuleInfoBlock(M);
 
+  // Output module level constants, used for global variable initializers
+  outputConstants(false);
+
   // Do the whole module now! Process each function at a time...
   for (Module::const_iterator I = M->begin(), E = M->end(); I != E; ++I)
-    processMethod(I);
+    outputFunction(I);
 
   // If needed, output the symbol table for the module...
   outputSymbolTable(M->getSymbolTable());
@@ -68,8 +85,9 @@ void BytecodeWriter::outputConstantsInPlane(const std::vector<const Value*>
                                             &Plane, unsigned StartNo) {
   unsigned ValNo = StartNo;
   
-  // Scan through and ignore function arguments...
-  for (; ValNo < Plane.size() && isa<Argument>(Plane[ValNo]); ValNo++)
+  // Scan through and ignore function arguments/global values...
+  for (; ValNo < Plane.size() && (isa<Argument>(Plane[ValNo]) ||
+                                  isa<GlobalValue>(Plane[ValNo])); ValNo++)
     /*empty*/;
 
   unsigned NC = ValNo;              // Number of constants
@@ -98,7 +116,7 @@ void BytecodeWriter::outputConstantsInPlane(const std::vector<const Value*>
       //     << Out.size() << "\n";
       outputConstant(CPV);
     } else {
-      outputType(cast<const Type>(V));
+      outputType(cast<Type>(V));
     }
   }
 }
@@ -108,26 +126,21 @@ void BytecodeWriter::outputConstants(bool isFunction) {
 
   unsigned NumPlanes = Table.getNumPlanes();
   
-  // Write the type plane for types first because earlier planes
-  // (e.g. for a primitive type like float) may have constants constructed
-  // using types coming later (e.g., via getelementptr from a pointer type).
-  // The type plane is needed before types can be fwd or bkwd referenced.
-  if (!isFunction) {
-    const std::vector<const Value*> &Plane = Table.getPlane(Type::TypeTyID);
-    assert(!Plane.empty() && "No types at all?");
-    unsigned ValNo = Type::FirstDerivedTyID; // Start at the derived types...
-    outputConstantsInPlane(Plane, ValNo);      // Write out the types
-  }
-  
   for (unsigned pno = 0; pno != NumPlanes; pno++) {
     const std::vector<const Value*> &Plane = Table.getPlane(pno);
-    if (!Plane.empty()) {             // Skip empty type planes...
+    if (!Plane.empty()) {              // Skip empty type planes...
       unsigned ValNo = 0;
-      if (isFunction)                   // Don't reemit module constants
-        ValNo = Table.getModuleLevel(pno);
+      if (isFunction)                 // Don't reemit module constants
+        ValNo += Table.getModuleLevel(pno);
       else if (pno == Type::TypeTyID) // If type plane wasn't written out above
         continue;
 
+      if (pno >= Type::FirstDerivedTyID) {
+        // Skip zero initializer
+        if (ValNo == 0)
+          ValNo = 1;
+      }
+
       outputConstantsInPlane(Plane, ValNo); // Write out constants in the plane
     }
   }
@@ -142,7 +155,7 @@ void BytecodeWriter::outputModuleInfoBlock(const Module *M) {
     assert(Slot != -1 && "Module global vars is broken!");
 
     // Fields: bit0 = isConstant, bit1 = hasInitializer, bit2=InternalLinkage,
-    // bit3+ = slot#
+    // bit3+ = Slot # for type
     unsigned oSlot = ((unsigned)Slot << 3) | (I->hasInternalLinkage() << 2) |
                      (I->hasInitializer() << 1) | I->isConstant();
     output_vbr(oSlot, Out);
@@ -165,11 +178,10 @@ void BytecodeWriter::outputModuleInfoBlock(const Module *M) {
   }
   output_vbr((unsigned)Table.getValSlot(Type::VoidTy), Out);
 
-
   align32(Out);
 }
 
-void BytecodeWriter::processMethod(const Function *F) {
+void BytecodeWriter::outputFunction(const Function *F) {
   BytecodeBlock FunctionBlock(BytecodeFormat::Function, Out);
   output_vbr((unsigned)F->hasInternalLinkage(), Out);
   // Only output the constant pool and other goodies if needed...

lib/Bytecode/Writer/WriterInternals.h

@@ -25,8 +25,8 @@ public:
   BytecodeWriter(std::deque<unsigned char> &o, const Module *M);
 
 protected:
-  void outputConstants(bool isMethod);
-  void processMethod(const Function *F);
+  void outputConstants(bool isFunction);
+  void outputFunction(const Function *F);
   void processBasicBlock(const BasicBlock &BB);
   void processInstruction(const Instruction &I);