Improve multi-threading and move Standalone to a multi-threading model (currently off though).

git-svn-id: https://cvs.khronos.org/svn/repos/ogl/trunk/ecosystem/public/sdk/tools/glslang@22565 e7fa87d3-cd2b-0410-9028-fcbf551c1848

glslang/MachineIndependent/Initialize.h

@@ -47,7 +47,7 @@ typedef TVector<TString> TBuiltInStrings;
 
 class TBuiltIns {
 public:
-    POOL_ALLOCATOR_NEW_DELETE(GlobalPoolAllocator)
+    POOL_ALLOCATOR_NEW_DELETE(GetThreadPoolAllocator())
     TBuiltIns();
     virtual ~TBuiltIns();
     void initialize(int version, EProfile);

glslang/MachineIndependent/PoolAlloc.cpp

@@ -46,7 +46,7 @@ void InitializeGlobalPools()
     if (globalPools)
         return;
 
-    TPoolAllocator *globalPoolAllocator = new TPoolAllocator(true);
+    TPoolAllocator *globalPoolAllocator = new TPoolAllocator();
 
     TThreadGlobalPools* threadData = new TThreadGlobalPools();
     
@@ -58,12 +58,12 @@ void InitializeGlobalPools()
 void FreeGlobalPools()
 {
     // Release the allocated memory for this thread.
-    TThreadGlobalPools* globalPools= static_cast<TThreadGlobalPools*>(OS_GetTLSValue(PoolIndex));    
-    if (!globalPools)
+    TThreadGlobalPools* globalPools = static_cast<TThreadGlobalPools*>(OS_GetTLSValue(PoolIndex));    
+    if (! globalPools)
         return;
 	
-    GlobalPoolAllocator.popAll();
-    delete &GlobalPoolAllocator;       
+    GetThreadPoolAllocator().popAll();
+    delete &GetThreadPoolAllocator();       
     delete globalPools;
 }
 
@@ -82,14 +82,14 @@ void FreePoolIndex()
     OS_FreeTLSIndex(PoolIndex);
 }
 
-TPoolAllocator& GetGlobalPoolAllocator()
+TPoolAllocator& GetThreadPoolAllocator()
 {
     TThreadGlobalPools* threadData = static_cast<TThreadGlobalPools*>(OS_GetTLSValue(PoolIndex));
 
     return *threadData->globalPoolAllocator;
 }
 
-void SetGlobalPoolAllocatorPtr(TPoolAllocator& poolAllocator)
+void SetThreadPoolAllocator(TPoolAllocator& poolAllocator)
 {
     TThreadGlobalPools* threadData = static_cast<TThreadGlobalPools*>(OS_GetTLSValue(PoolIndex));
 
@@ -100,8 +100,7 @@ void SetGlobalPoolAllocatorPtr(TPoolAllocator& poolAllocator)
 // Implement the functionality of the TPoolAllocator class, which
 // is documented in PoolAlloc.h.
 //
-TPoolAllocator::TPoolAllocator(bool g, int growthIncrement, int allocationAlignment) : 
-    global(g),
+TPoolAllocator::TPoolAllocator(int growthIncrement, int allocationAlignment) : 
     pageSize(growthIncrement),
     alignment(allocationAlignment),
     freeList(0),
@@ -148,19 +147,12 @@ TPoolAllocator::TPoolAllocator(bool g, int growthIncrement, int allocationAlignm
 
 TPoolAllocator::~TPoolAllocator()
 {
-    if (!global) {
-        //
-        // Then we know that this object is not being 
-        // allocated after other, globally scoped objects
-        // that depend on it.  So we can delete the "in use" memory.
-        //
-	    while (inUseList) {
-	        tHeader* next = inUseList->nextPage;
-            inUseList->~tHeader();
-            delete [] reinterpret_cast<char*>(inUseList);
-	        inUseList = next;
-	    }
-    }
+	while (inUseList) {
+	    tHeader* next = inUseList->nextPage;
+        inUseList->~tHeader();
+        delete [] reinterpret_cast<char*>(inUseList);
+	    inUseList = next;
+	}
 
     //
     // Always delete the free list memory - it can't be being

glslang/MachineIndependent/ShaderLang.cpp

@@ -79,7 +79,7 @@ int MapVersionToIndex(int version)
 const int VersionCount = 12;
 
 //
-// A symbol table per version per profile per language.  This will be sparsely
+// A process-global symbol table per version per profile per language.  This will be sparsely
 // populated, so they will only only be generated as needed.
 // 
 // Each has a different set of built-ins, and we want to preserve that from
@@ -163,41 +163,59 @@ bool AddContextSpecificSymbols(const TBuiltInResource* resources, TInfoSink& inf
     return true;
 }
 
+//
+// To do this on the fly, we want to leave the current state of our thread's 
+// pool allocator intact, so:
+//  - Switch to a new pool for parsing the built-ins
+//  - Do the parsing, which builds the symbol table, using the new pool
+//  - Switch to the process-global pool to save a copy the resulting symbol table
+//  - Free up the new pool used to parse the built-ins
+//  - Switch back to the original thread's pool
+//
+// This only gets done the first time any thread needs a particular symbol table
+// (lazy evaluation).
+//
 void SetupBuiltinSymbolTable(int version, EProfile profile)
 {
     TInfoSink infoSink;
 
-    // This function is for lazy setup.  See if already done.
+    // Make sure only one thread tries to do this at a time
+    glslang::GetGlobalLock();
+
+    // See if it's already been done.
     int versionIndex = MapVersionToIndex(version);
-    if (SharedSymbolTables[versionIndex][profile][EShLangVertex])
+    if (SharedSymbolTables[versionIndex][profile][EShLangVertex]) {
+        glslang::ReleaseGlobalLock();
+
         return;
+    }
 
-    TPoolAllocator& savedGPA = GetGlobalPoolAllocator();
-    TPoolAllocator *builtInPoolAllocator = new TPoolAllocator(true);
-    SetGlobalPoolAllocatorPtr(*builtInPoolAllocator);
+    // Switch to a new pool
+    TPoolAllocator& savedGPA = GetThreadPoolAllocator();
+    TPoolAllocator* builtInPoolAllocator = new TPoolAllocator();
+    SetThreadPoolAllocator(*builtInPoolAllocator);
 
+    // Generate the symbol table using the new pool
     TSymbolTable symTables[EShLangCount];
     if (profile == EEsProfile) {
         for (int stage = 0; stage < EShLangCount; ++stage)
             symTables[stage].setNoBuiltInRedeclarations();
     }
-
     GenerateBuiltInSymbolTable(infoSink, symTables, version, profile);
 
-    SetGlobalPoolAllocatorPtr(*PerProcessGPA);
+    // Switch to the process-global pool
+    SetThreadPoolAllocator(*PerProcessGPA);
 
+    // Copy the symbol table from the new pool to the process-global pool
     SharedSymbolTables[versionIndex][profile][EShLangVertex] = new TSymbolTable;
     SharedSymbolTables[versionIndex][profile][EShLangVertex]->copyTable(symTables[EShLangVertex]);
     SharedSymbolTables[versionIndex][profile][EShLangFragment] = new TSymbolTable;
     SharedSymbolTables[versionIndex][profile][EShLangFragment]->copyTable(symTables[EShLangFragment]);
-        
-    symTables[EShLangVertex].pop(0);
-    symTables[EShLangFragment].pop(0);
 
-    builtInPoolAllocator->popAll();
     delete builtInPoolAllocator;
+    SetThreadPoolAllocator(savedGPA);
 
-    SetGlobalPoolAllocatorPtr(savedGPA);
+    glslang::ReleaseGlobalLock();
 }
 
 bool DeduceProfile(TInfoSink& infoSink, int version, EProfile& profile)
@@ -261,7 +279,7 @@ int ShInitialize()
         return 0;
 
     if (! PerProcessGPA) { 
-        PerProcessGPA = new TPoolAllocator(true);
+        PerProcessGPA = new TPoolAllocator();
     }
     
     glslang::TScanContext::fillInKeywordMap();
@@ -333,6 +351,7 @@ int __fastcall ShFinalize()
         PerProcessGPA->popAll();
         delete PerProcessGPA;
     }
+
     return 1;
 }
 
@@ -361,6 +380,7 @@ int ShCompile(
 
     if (handle == 0)
         return 0;
+
     TShHandleBase* base = reinterpret_cast<TShHandleBase*>(handle);
     TCompiler* compiler = base->getAsCompiler();
     if (compiler == 0)
@@ -372,7 +392,7 @@ int ShCompile(
     if (numStrings == 0)
         return 1;
 
-    GlobalPoolAllocator.push();
+    GetThreadPoolAllocator().push();
     
     // move to length-based strings, rather than null-terminated strings
     int* lengths = new int[numStrings];
@@ -395,7 +415,6 @@ int ShCompile(
     bool goodProfile = DeduceProfile(compiler->infoSink, version, profile);
 
     TIntermediate intermediate(compiler->infoSink, version, profile);
-    
     SetupBuiltinSymbolTable(version, profile);
     TSymbolTable symbolTable(*SharedSymbolTables[MapVersionToIndex(version)]
                                                 [profile]
@@ -480,7 +499,7 @@ int ShCompile(
     //
     // Throw away all the temporary memory used by the compilation process.
     //
-    GlobalPoolAllocator.pop();
+    GetThreadPoolAllocator().pop();
     delete [] lengths;
 
     return success ? 1 : 0;
@@ -510,9 +529,9 @@ int ShLink(
         return 0;
 
     int returnValue;
-    GlobalPoolAllocator.push();
+    GetThreadPoolAllocator().push();
     returnValue = ShLinkExt(linkHandle, compHandles, numHandles);
-    GlobalPoolAllocator.pop();
+    GetThreadPoolAllocator().pop();
 
     if (returnValue)
         return 1;

glslang/MachineIndependent/SymbolTable.h

@@ -77,7 +77,7 @@ class TFunction;
 class TAnonMember;
 class TSymbol {
 public:
-    POOL_ALLOCATOR_NEW_DELETE(GlobalPoolAllocator)
+    POOL_ALLOCATOR_NEW_DELETE(GetThreadPoolAllocator())
     explicit TSymbol(const TString *n) :  name(n) { }
 	virtual TSymbol* clone(TStructureMap& remapper) = 0;
     virtual ~TSymbol() { }
@@ -245,7 +245,7 @@ protected:
 
 class TSymbolTableLevel {
 public:
-    POOL_ALLOCATOR_NEW_DELETE(GlobalPoolAllocator)
+    POOL_ALLOCATOR_NEW_DELETE(GetThreadPoolAllocator())
     TSymbolTableLevel() : defaultPrecision (0), anonId(0) { }
 	~TSymbolTableLevel();
 
@@ -365,24 +365,23 @@ protected:
 
 class TSymbolTable {
 public:
-    TSymbolTable() : uniqueId(0), noBuiltInRedeclarations(false)
+    TSymbolTable() : uniqueId(0), noBuiltInRedeclarations(false), adoptedLevels(1)  // TODO: memory: can we make adoptedLevels be 0 for symbol tables we don't keep?
     {
         //
-        // The symbol table cannot be used until push() is called, but
-        // the lack of an initial call to push() can be used to detect
-        // that the symbol table has not been preloaded with built-ins.
+        // This symbol table cannot be used until push() is called.
         //
     }
     explicit TSymbolTable(TSymbolTable& symTable)
     {
         table.push_back(symTable.table[0]);
+        adoptedLevels = 1;
         uniqueId = symTable.uniqueId;
         noBuiltInRedeclarations = symTable.noBuiltInRedeclarations;
     }
     ~TSymbolTable()
     {
-        // level 0 is always built-in symbols, so we never pop that out
-        while (table.size() > 1)
+        // don't deallocate levels passed in from elsewhere
+        while (table.size() > adoptedLevels)
             pop(0);
     }
 
@@ -463,6 +462,7 @@ protected:
     std::vector<TSymbolTableLevel*> table;
     int uniqueId;     // for unique identification in code generation
     bool noBuiltInRedeclarations;
+    unsigned int adoptedLevels;
 };
 
 #endif // _SYMBOL_TABLE_INCLUDED_

glslang/MachineIndependent/glslang.y

@@ -319,6 +319,7 @@ postfix_expression
             TType newType($1->getType());
             newType.dereference();
             $$->setType(newType);
+            // TODO: functionality: does this drop const qualification for const[const] ?
         }
     }
     | function_call {

glslang/MachineIndependent/localintermediate.h

@@ -51,7 +51,7 @@ struct TVectorFields {
 class TInfoSink;
 class TIntermediate {
 public:    
-    POOL_ALLOCATOR_NEW_DELETE(GlobalPoolAllocator)
+    POOL_ALLOCATOR_NEW_DELETE(GetThreadPoolAllocator())
 
     TIntermediate(TInfoSink& i, int v, EProfile p) : infoSink(i), version(v), profile(p) { }
     TIntermSymbol* addSymbol(int Id, const TString&, const TType&, TSourceLoc);