Ok, I can't handle it. This is a temporary checkin of a ton of statistics that

i'm using in my work to reduce the bytecode file sizes.  These will eventually
be removed.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@10849 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 100 insertions, 9 deletions

diff --git a/lib/Bytecode/Writer/ConstantWriter.cpp b/lib/Bytecode/Writer/ConstantWriter.cpp
index 30d93c9..12043b0 100644
--- a/lib/Bytecode/Writer/ConstantWriter.cpp
+++ b/lib/Bytecode/Writer/ConstantWriter.cpp
@@ -16,15 +16,26 @@
 #include "llvm/Constants.h"
 #include "llvm/SymbolTable.h"
 #include "llvm/DerivedTypes.h"
+#include "Support/Statistic.h"
 using namespace llvm;
 
+static Statistic<> 
+TypeBytes("bytecodewriter", "Bytes of types");
+static Statistic<> 
+ConstantBytes("bytecodewriter", "Bytes of constants");
+static Statistic<> 
+NumConstants("bytecodewriter", "Number of constants");
+
 void BytecodeWriter::outputType(const Type *T) {
+  TypeBytes -= Out.size();
   output_vbr((unsigned)T->getPrimitiveID(), Out);
   
   // That's all there is to handling primitive types...
-  if (T->isPrimitiveType())
+  if (T->isPrimitiveType()) {
+    TypeBytes += Out.size();
     return;     // We might do this if we alias a prim type: %x = type int
-  
+  }
+
   switch (T->getPrimitiveID()) {   // Handle derived types now.
   case Type::FunctionTyID: {
     const FunctionType *MT = cast<FunctionType>(T);
@@ -95,12 +106,16 @@ void BytecodeWriter::outputType(const Type *T) {
               << " Type '" << T->getDescription() << "'\n";
     break;
   }
+  TypeBytes += Out.size();
 }
 
 bool BytecodeWriter::outputConstant(const Constant *CPV) {
+  ConstantBytes -= Out.size();
   assert((CPV->getType()->isPrimitiveType() || !CPV->isNullValue()) &&
          "Shouldn't output null constants!");
 
+  ++NumConstants;
+
   // We must check for a ConstantExpr before switching by type because
   // a ConstantExpr can be of any type, and has no explicit value.
   // 
@@ -117,9 +132,10 @@ bool BytecodeWriter::outputConstant(const Constant *CPV) {
       Slot = Table.getSlot((*OI)->getType());
       output_vbr((unsigned)Slot, Out);
     }
+    ConstantBytes += Out.size();
     return false;
   } else {
-    output_vbr((unsigned)0, Out);       // flag as not a ConstantExpr
+    output_vbr(0U, Out);       // flag as not a ConstantExpr
   }
   
   switch (CPV->getType()->getPrimitiveID()) {
@@ -169,7 +185,7 @@ bool BytecodeWriter::outputConstant(const Constant *CPV) {
       int Slot = Table.getSlot(Vals[i]);
       assert(Slot != -1 && "Constant used but not available!!");
       output_vbr((unsigned)Slot, Out);
-    }      
+    }
     break;
   }
 
@@ -199,5 +215,6 @@ bool BytecodeWriter::outputConstant(const Constant *CPV) {
               << " type '" << CPV->getType()->getName() << "'\n";
     break;
   }
+    ConstantBytes += Out.size();
   return false;
 }
diff --git a/lib/Bytecode/Writer/InstructionWriter.cpp b/lib/Bytecode/Writer/InstructionWriter.cpp
index bc5268c..0bf555d 100644
--- a/lib/Bytecode/Writer/InstructionWriter.cpp
+++ b/lib/Bytecode/Writer/InstructionWriter.cpp
@@ -22,6 +22,23 @@ using namespace llvm;
 
 static Statistic<> 
 NumInstrs("bytecodewriter", "Number of instructions");
+static Statistic<> 
+NumOversizedInstrs("bytecodewriter", "Number of oversized instructions");
+static Statistic<> 
+BytesOversizedInstrs("bytecodewriter", "Bytes of oversized instructions");
+
+static Statistic<> 
+NumHugeOperandInstrs("bytecodewriter", "Number of instructions with > 3 operands");
+static Statistic<> 
+NumOversized1OpInstrs("bytecodewriter", "Number of oversized 1 operand instrs");
+static Statistic<> 
+NumOversized2OpInstrs("bytecodewriter", "Number of oversized 2 operand instrs");
+static Statistic<>
+NumOversized3OpInstrs("bytecodewriter", "Number of oversized 3 operand instrs");
+
+static Statistic<>
+NumOversidedBecauseOfTypes("bytecodewriter", "Number of oversized instructions because of their type");
+
 
 typedef unsigned char uchar;
 
@@ -33,6 +50,9 @@ typedef unsigned char uchar;
 static void outputInstructionFormat0(const Instruction *I, unsigned Opcode,
 				     const SlotCalculator &Table,
 				     unsigned Type, std::deque<uchar> &Out) {
+  NumOversizedInstrs++;
+  BytesOversizedInstrs -= Out.size();
+
   // Opcode must have top two bits clear...
   output_vbr(Opcode << 2, Out);                  // Instruction Opcode ID
   output_vbr(Type, Out);                         // Result type
@@ -58,6 +78,7 @@ static void outputInstructionFormat0(const Instruction *I, unsigned Opcode,
   }
 
   align32(Out);    // We must maintain correct alignment!
+  BytesOversizedInstrs += Out.size();
 }
 
 
@@ -275,6 +296,10 @@ void BytecodeWriter::processInstruction(const Instruction &I) {
       outputInstructionFormat1(&I, Opcode, Table, Slots, Type, Out);
       return;
     }
+    if (Type >= (1 << 12)-1)
+      NumOversidedBecauseOfTypes++;
+
+    NumOversized1OpInstrs++;
     break;
 
   case 2:
@@ -282,6 +307,9 @@ void BytecodeWriter::processInstruction(const Instruction &I) {
       outputInstructionFormat2(&I, Opcode, Table, Slots, Type, Out);
       return;
     }
+    if (Type >= (1 << 8)) 
+      NumOversidedBecauseOfTypes++;
+    NumOversized2OpInstrs++;
     break;
 
   case 3:
@@ -289,6 +317,12 @@ void BytecodeWriter::processInstruction(const Instruction &I) {
       outputInstructionFormat3(&I, Opcode, Table, Slots, Type, Out);
       return;
     }
+    if (Type >= (1 << 6)) 
+      NumOversidedBecauseOfTypes++;
+    NumOversized3OpInstrs++;
+    break;
+  default:
+    ++NumHugeOperandInstrs;
     break;
   }
 
diff --git a/lib/Bytecode/Writer/Writer.cpp b/lib/Bytecode/Writer/Writer.cpp
index 100d34e..ca50b5a 100644
--- a/lib/Bytecode/Writer/Writer.cpp
+++ b/lib/Bytecode/Writer/Writer.cpp
@@ -29,6 +29,7 @@
 #include "llvm/DerivedTypes.h"
 #include "Support/STLExtras.h"
 #include "Support/Statistic.h"
+#include "Support/Debug.h"
 #include <cstring>
 #include <algorithm>
 using namespace llvm;
@@ -37,7 +38,18 @@ static RegisterPass<WriteBytecodePass> X("emitbytecode", "Bytecode Writer");
 
 static Statistic<> 
 BytesWritten("bytecodewriter", "Number of bytecode bytes written");
-
+static Statistic<> 
+ConstantTotalBytes("bytecodewriter", "Bytes of constants total");
+static Statistic<> 
+FunctionConstantTotalBytes("bytecodewriter", "Bytes of function constants total");
+static Statistic<>
+ConstantPlaneHeaderBytes("bytecodewriter", "Constant plane header bytes");
+static Statistic<> 
+InstructionBytes("bytecodewriter", "Bytes of bytes of instructions");
+static Statistic<> 
+SymTabBytes("bytecodewriter", "Bytes of symbol table");
+static Statistic<> 
+ModuleInfoBytes("bytecodewriter", "Bytes of module info");
 
 BytecodeWriter::BytecodeWriter(std::deque<unsigned char> &o, const Module *M) 
   : Out(o), Table(M, true) {
@@ -52,8 +64,9 @@ BytecodeWriter::BytecodeWriter(std::deque<unsigned char> &o, const Module *M)
   bool hasNoEndianness  = M->getEndianness() == Module::AnyEndianness;
   bool hasNoPointerSize = M->getPointerSize() == Module::AnyPointerSize;
 
-  // Output the version identifier... we are currently on bytecode version #0
-  unsigned Version = (0 << 4) | isBigEndian | (hasLongPointers << 1) |
+  // Output the version identifier... we are currently on bytecode version #1,
+  // which corresponds to LLVM v1.2.
+  unsigned Version = (1 << 4) | isBigEndian | (hasLongPointers << 1) |
                      (hasNoEndianness << 2) | (hasNoPointerSize << 3);
   output_vbr(Version, Out);
   align32(Out);
@@ -71,6 +84,12 @@ BytecodeWriter::BytecodeWriter(std::deque<unsigned char> &o, const Module *M)
     outputConstantsInPlane(Plane, ValNo);      // Write out the types
   }
 
+  DEBUG(for (unsigned i = 0; i != Type::TypeTyID; ++i)
+          if (Table.getPlane(i).size())
+            std::cerr << "  ModuleLevel["
+                      << *Type::getPrimitiveType((Type::PrimitiveID)i)
+                      << "] = " << Table.getPlane(i).size() << "\n");
+
   // The ModuleInfoBlock follows directly after the type information
   outputModuleInfoBlock(M);
 
@@ -104,6 +123,11 @@ void BytecodeWriter::outputConstantsInPlane(const std::vector<const Value*>
   NC -= ValNo;                      // Convert from index into count
   if (NC == 0) return;              // Skip empty type planes...
 
+  // FIXME: Most slabs only have 1 or 2 entries!  We should encode this much
+  // more compactly.
+
+  ConstantPlaneHeaderBytes -= Out.size();
+
   // Output type header: [num entries][type id number]
   //
   output_vbr(NC, Out);
@@ -113,6 +137,9 @@ void BytecodeWriter::outputConstantsInPlane(const std::vector<const Value*>
   assert (Slot != -1 && "Type in constant pool but not in function!!");
   output_vbr((unsigned)Slot, Out);
 
+  ConstantPlaneHeaderBytes += Out.size();
+
+
   //cerr << "Emitting " << NC << " constants of type '" 
   //	 << Plane.front()->getType()->getName() << "' = Slot #" << Slot << "\n";
 
@@ -129,6 +156,8 @@ void BytecodeWriter::outputConstantsInPlane(const std::vector<const Value*>
 }
 
 void BytecodeWriter::outputConstants(bool isFunction) {
+  ConstantTotalBytes -= Out.size();
+  if (isFunction) FunctionConstantTotalBytes -= Out.size();
   BytecodeBlock CPool(BytecodeFormat::ConstantPool, Out);
 
   unsigned NumPlanes = Table.getNumPlanes();
@@ -160,6 +189,8 @@ void BytecodeWriter::outputConstants(bool isFunction) {
         outputConstantsInPlane(Plane, ValNo);
       }
     }
+  ConstantTotalBytes += Out.size();
+  if (isFunction) FunctionConstantTotalBytes += Out.size();
 }
 
 static unsigned getEncodedLinkage(const GlobalValue *GV) {
@@ -174,6 +205,8 @@ static unsigned getEncodedLinkage(const GlobalValue *GV) {
 }
 
 void BytecodeWriter::outputModuleInfoBlock(const Module *M) {
+  ModuleInfoBytes -= Out.size();
+
   BytecodeBlock ModuleInfoBlock(BytecodeFormat::ModuleGlobalInfo, Out);
   
   // Output the types for the global variables in the module...
@@ -206,6 +239,8 @@ void BytecodeWriter::outputModuleInfoBlock(const Module *M) {
   output_vbr((unsigned)Table.getSlot(Type::VoidTy), Out);
 
   align32(Out);
+
+  ModuleInfoBytes += Out.size();
 }
 
 void BytecodeWriter::outputFunction(const Function *F) {
@@ -222,10 +257,11 @@ void BytecodeWriter::outputFunction(const Function *F) {
 
     {  // Output all of the instructions in the body of the function
       BytecodeBlock ILBlock(BytecodeFormat::InstructionList, Out);
-
+      InstructionBytes -= Out.size();
       for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E;++BB)
         for(BasicBlock::const_iterator I = BB->begin(), E = BB->end(); I!=E;++I)
           processInstruction(*I);
+      InstructionBytes += Out.size();
     }
     
     // If needed, output the symbol table for the function...
@@ -240,7 +276,9 @@ void BytecodeWriter::outputSymbolTable(const SymbolTable &MST) {
   // space!
   if (MST.begin() == MST.end()) return;
 
-  BytecodeBlock FunctionBlock(BytecodeFormat::SymbolTable, Out);
+  SymTabBytes -= Out.size();
+  
+  BytecodeBlock SymTabBlock(BytecodeFormat::SymbolTable, Out);
 
   for (SymbolTable::const_iterator TI = MST.begin(); TI != MST.end(); ++TI) {
     SymbolTable::type_const_iterator I = MST.type_begin(TI->first);
@@ -264,6 +302,8 @@ void BytecodeWriter::outputSymbolTable(const SymbolTable &MST) {
       output(I->first, Out, false); // Don't force alignment...
     }
   }
+
+  SymTabBytes += Out.size();
 }
 
 void llvm::WriteBytecodeToFile(const Module *C, std::ostream &Out) {