Merge pull request #1147 from KhronosGroup/memory1

Memory fixes, round 1

OGLCompilersDLL/InitializeDll.cpp

@@ -38,13 +38,17 @@
 
 #include "InitializeDll.h"
 #include "../glslang/Include/InitializeGlobals.h"
-
 #include "../glslang/Public/ShaderLang.h"
+#include "../glslang/Include/PoolAlloc.h"
 
 namespace glslang {
 
 OS_TLSIndex ThreadInitializeIndex = OS_INVALID_TLS_INDEX;
 
+// Per-process initialization.
+// Needs to be called at least once before parsing, etc. is done.
+// Will also do thread initialization for the calling thread; other
+// threads will need to do that explicitly.
 bool InitProcess()
 {
     glslang::GetGlobalLock();
@@ -85,7 +89,9 @@ bool InitProcess()
     return true;
 }
 
-
+// Per-thread scoped initialization.
+// Must be called at least once by each new thread sharing the
+// symbol tables, etc., needed to parse.
 bool InitThread()
 {
     //
@@ -99,17 +105,21 @@ bool InitThread()
     if (OS_GetTLSValue(ThreadInitializeIndex) != 0)
         return true;
 
-    InitializeMemoryPools();
-
     if (! OS_SetTLSValue(ThreadInitializeIndex, (void *)1)) {
         assert(0 && "InitThread(): Unable to set init flag.");
         return false;
     }
 
+    glslang::SetThreadPoolAllocator(nullptr);
+
     return true;
 }
 
-
+// Not necessary to call this: InitThread() is reentrant, and the need
+// to do per thread tear down has been removed.
+//
+// This is kept, with memory management removed, to satisfy any exiting
+// calls to it that rely on it.
 bool DetachThread()
 {
     bool success = true;
@@ -125,14 +135,18 @@ bool DetachThread()
             assert(0 && "DetachThread(): Unable to clear init flag.");
             success = false;
         }
-
-        FreeGlobalPools();
-
     }
 
     return success;
 }
 
+// Not necessary to call this: InitProcess() is reentrant.
+//
+// This is kept, with memory management removed, to satisfy any exiting
+// calls to it that rely on it.
+//
+// Users of glslang should call shFinalize() or glslang::FinalizeProcess() for
+// process-scoped memory tear down.
 bool DetachProcess()
 {
     bool success = true;
@@ -140,12 +154,8 @@ bool DetachProcess()
     if (ThreadInitializeIndex == OS_INVALID_TLS_INDEX)
         return true;
 
-    ShFinalize();
-
     success = DetachThread();
 
-    FreePoolIndex();
-
     OS_FreeTLSIndex(ThreadInitializeIndex);
     ThreadInitializeIndex = OS_INVALID_TLS_INDEX;
 

OGLCompilersDLL/InitializeDll.h

@@ -40,8 +40,8 @@ namespace glslang {
 
 bool InitProcess();
 bool InitThread();
-bool DetachThread();
-bool DetachProcess();
+bool DetachThread();  // not called from standalone, perhaps other tools rely on parts of it
+bool DetachProcess(); // not called from standalone, perhaps other tools rely on parts of it
 
 } // end namespace glslang
 

StandAlone/StandAlone.cpp

@@ -127,6 +127,9 @@ void InfoLogMsg(const char* msg, const char* name, const int num);
 bool CompileFailed = false;
 bool LinkFailed = false;
 
+// array of unique places to leave the shader names and infologs for the asynchronous compiles
+std::vector<std::unique_ptr<glslang::TWorkItem>> WorkItems;
+
 TBuiltInResource Resources;
 std::string ConfigFile;
 
@@ -1022,14 +1025,10 @@ void CompileAndLinkShaderFiles(glslang::TWorklist& Worklist)
         FreeFileData(const_cast<char*>(it->text[0]));
 }
 
-int C_DECL main(int argc, char* argv[])
+int singleMain()
 {
-    // array of unique places to leave the shader names and infologs for the asynchronous compiles
-    std::vector<std::unique_ptr<glslang::TWorkItem>> workItems;
-    ProcessArguments(workItems, argc, argv);
-
     glslang::TWorklist workList;
-    std::for_each(workItems.begin(), workItems.end(), [&workList](std::unique_ptr<glslang::TWorkItem>& item) {
+    std::for_each(WorkItems.begin(), WorkItems.end(), [&workList](std::unique_ptr<glslang::TWorkItem>& item) {
         assert(item);
         workList.add(item.get());
     });
@@ -1061,8 +1060,8 @@ int C_DECL main(int argc, char* argv[])
     }
 
     if (Options & EOptionStdin) {
-        workItems.push_back(std::unique_ptr<glslang::TWorkItem>{new glslang::TWorkItem("stdin")});
-        workList.add(workItems.back().get());
+        WorkItems.push_back(std::unique_ptr<glslang::TWorkItem>{new glslang::TWorkItem("stdin")});
+        workList.add(WorkItems.back().get());
     }
 
     ProcessConfigFile();
@@ -1100,11 +1099,11 @@ int C_DECL main(int argc, char* argv[])
             CompileShaders(workList);
 
         // Print out all the resulting infologs
-        for (size_t w = 0; w < workItems.size(); ++w) {
-            if (workItems[w]) {
-                if (printShaderNames || workItems[w]->results.size() > 0)
-                    PutsIfNonEmpty(workItems[w]->name.c_str());
-                PutsIfNonEmpty(workItems[w]->results.c_str());
+        for (size_t w = 0; w < WorkItems.size(); ++w) {
+            if (WorkItems[w]) {
+                if (printShaderNames || WorkItems[w]->results.size() > 0)
+                    PutsIfNonEmpty(WorkItems[w]->name.c_str());
+                PutsIfNonEmpty(WorkItems[w]->results.c_str());
             }
         }
 
@@ -1119,6 +1118,25 @@ int C_DECL main(int argc, char* argv[])
     return 0;
 }
 
+int C_DECL main(int argc, char* argv[])
+{
+    ProcessArguments(WorkItems, argc, argv);
+
+    int ret = 0;
+
+    // Loop over the entire init/finalize cycle to watch memory changes
+    const int iterations = 1;
+    if (iterations > 1)
+        glslang::OS_DumpMemoryCounters();
+    for (int i = 0; i < iterations; ++i) {
+        ret = singleMain();
+        if (iterations > 1)
+            glslang::OS_DumpMemoryCounters();
+    }
+
+    return ret;
+}
+
 //
 //   Deduce the language from the filename.  Files must end in one of the
 //   following extensions:

glslang/Include/InitializeGlobals.h

@@ -37,10 +37,7 @@
 
 namespace glslang {
 
-void InitializeMemoryPools();
-void FreeGlobalPools();
 bool InitializePoolIndex();
-void FreePoolIndex();
 
 } // end namespace glslang
 

glslang/Include/PoolAlloc.h

@@ -250,15 +250,8 @@ private:
 // different times.  But a simple use is to have a global pop
 // with everyone using the same global allocator.
 //
-typedef TPoolAllocator* PoolAllocatorPointer;
 extern TPoolAllocator& GetThreadPoolAllocator();
-
-struct TThreadMemoryPools
-{
-    TPoolAllocator* threadPoolAllocator;
-};
-
-void SetThreadPoolAllocator(TPoolAllocator& poolAllocator);
+void SetThreadPoolAllocator(TPoolAllocator* poolAllocator);
 
 //
 // This STL compatible allocator is intended to be used as the allocator

glslang/Include/ShHandle.h

@@ -56,11 +56,14 @@ class TUniformMap;
 //
 class TShHandleBase {
 public:
-    TShHandleBase() { }
-    virtual ~TShHandleBase() { }
+    TShHandleBase() { pool = new glslang::TPoolAllocator; }
+    virtual ~TShHandleBase() { delete pool; }
     virtual TCompiler* getAsCompiler() { return 0; }
     virtual TLinker* getAsLinker() { return 0; }
     virtual TUniformMap* getAsUniformMap() { return 0; }
+    virtual glslang::TPoolAllocator* getPool() const { return pool; }
+private:
+    glslang::TPoolAllocator* pool;
 };
 
 //

glslang/MachineIndependent/PoolAlloc.cpp

@@ -40,35 +40,22 @@
 
 namespace glslang {
 
+// Process-wide TLS index
 OS_TLSIndex PoolIndex;
 
-void InitializeMemoryPools()
+// Return the thread-specific current pool.
+TPoolAllocator& GetThreadPoolAllocator()
 {
-    TThreadMemoryPools* pools = static_cast<TThreadMemoryPools*>(OS_GetTLSValue(PoolIndex));
-    if (pools)
-        return;
-
-    TPoolAllocator *threadPoolAllocator = new TPoolAllocator();
-
-    TThreadMemoryPools* threadData = new TThreadMemoryPools();
-
-    threadData->threadPoolAllocator = threadPoolAllocator;
-
-    OS_SetTLSValue(PoolIndex, threadData);
+    return *static_cast<TPoolAllocator*>(OS_GetTLSValue(PoolIndex));
 }
 
-void FreeGlobalPools()
+// Set the thread-specific current pool.
+void SetThreadPoolAllocator(TPoolAllocator* poolAllocator)
 {
-    // Release the allocated memory for this thread.
-    TThreadMemoryPools* globalPools = static_cast<TThreadMemoryPools*>(OS_GetTLSValue(PoolIndex));
-    if (! globalPools)
-        return;
-
-    GetThreadPoolAllocator().popAll();
-    delete &GetThreadPoolAllocator();
-    delete globalPools;
+    OS_SetTLSValue(PoolIndex, poolAllocator);
 }
 
+// Process-wide set up of the TLS pool storage.
 bool InitializePoolIndex()
 {
     // Allocate a TLS index.
@@ -78,26 +65,6 @@ bool InitializePoolIndex()
     return true;
 }
 
-void FreePoolIndex()
-{
-    // Release the TLS index.
-    OS_FreeTLSIndex(PoolIndex);
-}
-
-TPoolAllocator& GetThreadPoolAllocator()
-{
-    TThreadMemoryPools* threadData = static_cast<TThreadMemoryPools*>(OS_GetTLSValue(PoolIndex));
-
-    return *threadData->threadPoolAllocator;
-}
-
-void SetThreadPoolAllocator(TPoolAllocator& poolAllocator)
-{
-    TThreadMemoryPools* threadData = static_cast<TThreadMemoryPools*>(OS_GetTLSValue(PoolIndex));
-
-    threadData->threadPoolAllocator = &poolAllocator;
-}
-
 //
 // Implement the functionality of the TPoolAllocator class, which
 // is documented in PoolAlloc.h.

glslang/MachineIndependent/ShaderLang.cpp

@@ -217,7 +217,7 @@ enum EPrecisionClass {
 TSymbolTable* CommonSymbolTable[VersionCount][SpvVersionCount][ProfileCount][SourceCount][EPcCount] = {};
 TSymbolTable* SharedSymbolTables[VersionCount][SpvVersionCount][ProfileCount][SourceCount][EShLangCount] = {};
 
-TPoolAllocator* PerProcessGPA = 0;
+TPoolAllocator* PerProcessGPA = nullptr;
 
 //
 // Parse and add to the given symbol table the content of the given shader string.
@@ -361,7 +361,7 @@ bool AddContextSpecificSymbols(const TBuiltInResource* resources, TInfoSink& inf
 // 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
+//  - Switch to the process-global pool to save a copy of the resulting symbol table
 //  - Free up the new pool used to parse the built-ins
 //  - Switch back to the original thread's pool
 //
@@ -388,8 +388,8 @@ void SetupBuiltinSymbolTable(int version, EProfile profile, const SpvVersion& sp
 
     // Switch to a new pool
     TPoolAllocator& previousAllocator = GetThreadPoolAllocator();
-    TPoolAllocator* builtInPoolAllocator = new TPoolAllocator();
-    SetThreadPoolAllocator(*builtInPoolAllocator);
+    TPoolAllocator* builtInPoolAllocator = new TPoolAllocator;
+    SetThreadPoolAllocator(builtInPoolAllocator);
 
     // Dynamically allocate the local symbol tables so we can control when they are deallocated WRT when the pool is popped.
     TSymbolTable* commonTable[EPcCount];
@@ -403,7 +403,7 @@ void SetupBuiltinSymbolTable(int version, EProfile profile, const SpvVersion& sp
     InitializeSymbolTables(infoSink, commonTable, stageTables, version, profile, spvVersion, source);
 
     // Switch to the process-global pool
-    SetThreadPoolAllocator(*PerProcessGPA);
+    SetThreadPoolAllocator(PerProcessGPA);
 
     // Copy the local symbol tables from the new pool to the global tables using the process-global pool
     for (int precClass = 0; precClass < EPcCount; ++precClass) {
@@ -430,7 +430,7 @@ void SetupBuiltinSymbolTable(int version, EProfile profile, const SpvVersion& sp
         delete stageTables[stage];
 
     delete builtInPoolAllocator;
-    SetThreadPoolAllocator(previousAllocator);
+    SetThreadPoolAllocator(&previousAllocator);
 
     glslang::ReleaseGlobalLock();
 }
@@ -722,9 +722,6 @@ bool ProcessDeferred(
     const std::string sourceEntryPointName = "",
     const TEnvironment* environment = nullptr)  // optional way of fully setting all versions, overriding the above
 {
-    if (! InitThread())
-        return false;
-
     // This must be undone (.pop()) by the caller, after it finishes consuming the created tree.
     GetThreadPoolAllocator().push();
 
@@ -1196,7 +1193,7 @@ int ShInitialize()
     if (! InitProcess())
         return 0;
 
-    if (! PerProcessGPA)
+    if (PerProcessGPA == nullptr)
         PerProcessGPA = new TPoolAllocator();
 
     glslang::TScanContext::fillInKeywordMap();
@@ -1288,10 +1285,9 @@ int __fastcall ShFinalize()
         }
     }
 
-    if (PerProcessGPA) {
-        PerProcessGPA->popAll();
+    if (PerProcessGPA != nullptr) {
         delete PerProcessGPA;
-        PerProcessGPA = 0;
+        PerProcessGPA = nullptr;
     }
 
     glslang::TScanContext::deleteKeywordMap();
@@ -1332,6 +1328,8 @@ int ShCompile(
     if (compiler == 0)
         return 0;
 
+    SetThreadPoolAllocator(compiler->getPool());
+
     compiler->infoSink.info.erase();
     compiler->infoSink.debug.erase();
 
@@ -1389,6 +1387,8 @@ int ShLinkExt(
     TShHandleBase* base = reinterpret_cast<TShHandleBase*>(linkHandle);
     TLinker* linker = static_cast<TLinker*>(base->getAsLinker());
 
+    SetThreadPoolAllocator(linker->getPool());
+
     if (linker == 0)
         return 0;
 
@@ -1423,9 +1423,6 @@ void ShSetEncryptionMethod(ShHandle handle)
 //
 const char* ShGetInfoLog(const ShHandle handle)
 {
-    if (!InitThread())
-        return 0;
-
     if (handle == 0)
         return 0;
 
@@ -1449,9 +1446,6 @@ const char* ShGetInfoLog(const ShHandle handle)
 //
 const void* ShGetExecutable(const ShHandle handle)
 {
-    if (!InitThread())
-        return 0;
-
     if (handle == 0)
         return 0;
 
@@ -1474,9 +1468,6 @@ const void* ShGetExecutable(const ShHandle handle)
 //
 int ShSetVirtualAttributeBindings(const ShHandle handle, const ShBindingTable* table)
 {
-    if (!InitThread())
-        return 0;
-
     if (handle == 0)
         return 0;
 
@@ -1496,9 +1487,6 @@ int ShSetVirtualAttributeBindings(const ShHandle handle, const ShBindingTable* t
 //
 int ShSetFixedAttributeBindings(const ShHandle handle, const ShBindingTable* table)
 {
-    if (!InitThread())
-        return 0;
-
     if (handle == 0)
         return 0;
 
@@ -1517,9 +1505,6 @@ int ShSetFixedAttributeBindings(const ShHandle handle, const ShBindingTable* tab
 //
 int ShExcludeAttributes(const ShHandle handle, int *attributes, int count)
 {
-    if (!InitThread())
-        return 0;
-
     if (handle == 0)
         return 0;
 
@@ -1541,9 +1526,6 @@ int ShExcludeAttributes(const ShHandle handle, int *attributes, int count)
 //
 int ShGetUniformLocation(const ShHandle handle, const char* name)
 {
-    if (!InitThread())
-        return 0;
-
     if (handle == 0)
         return -1;
 
@@ -1602,8 +1584,9 @@ public:
 };
 
 TShader::TShader(EShLanguage s)
-    : pool(0), stage(s), lengths(nullptr), stringNames(nullptr), preamble("")
+    : stage(s), lengths(nullptr), stringNames(nullptr), preamble("")
 {
+    pool = new TPoolAllocator;
     infoSink = new TInfoSink;
     compiler = new TDeferredCompiler(stage, *infoSink);
     intermediate = new TIntermediate(s);
@@ -1706,9 +1689,8 @@ bool TShader::parse(const TBuiltInResource* builtInResources, int defaultVersion
 {
     if (! InitThread())
         return false;
+    SetThreadPoolAllocator(pool);
 
-    pool = new TPoolAllocator();
-    SetThreadPoolAllocator(*pool);
     if (! preamble)
         preamble = "";
 
@@ -1730,9 +1712,8 @@ bool TShader::preprocess(const TBuiltInResource* builtInResources,
 {
     if (! InitThread())
         return false;
+    SetThreadPoolAllocator(pool);
 
-    pool = new TPoolAllocator();
-    SetThreadPoolAllocator(*pool);
     if (! preamble)
         preamble = "";
 
@@ -1752,8 +1733,9 @@ const char* TShader::getInfoDebugLog()
     return infoSink->debug.c_str();
 }
 
-TProgram::TProgram() : pool(0), reflection(0), ioMapper(nullptr), linked(false)
+TProgram::TProgram() : reflection(0), ioMapper(nullptr), linked(false)
 {
+    pool = new TPoolAllocator;
     infoSink = new TInfoSink;
     for (int s = 0; s < EShLangCount; ++s) {
         intermediate[s] = 0;
@@ -1788,8 +1770,7 @@ bool TProgram::link(EShMessages messages)
 
     bool error = false;
 
-    pool = new TPoolAllocator();
-    SetThreadPoolAllocator(*pool);
+    SetThreadPoolAllocator(pool);
 
     for (int s = 0; s < EShLangCount; ++s) {
         if (! linkStage((EShLanguage)s, messages))

glslang/OSDependent/Unix/ossource.cpp

@@ -43,6 +43,9 @@
 #include <assert.h>
 #include <errno.h>
 #include <stdint.h>
+#include <cstdio>
+#include <sys/time.h>
+#include <sys/resource.h>
 
 namespace glslang {
 
@@ -184,8 +187,18 @@ void ReleaseGlobalLock()
   pthread_mutex_unlock(&gMutex);
 }
 
+// #define DUMP_COUNTERS
+
 void OS_DumpMemoryCounters()
 {
+#ifdef DUMP_COUNTERS
+    struct rusage usage;
+
+    if (getrusage(RUSAGE_SELF, &usage) == 0)
+        printf("Working set size: %ld\n", usage.ru_maxrss * 1024);
+#else
+    printf("Recompile with DUMP_COUNTERS defined to see counters.\n");
+#endif
 }
 
 } // end namespace glslang

glslang/Public/ShaderLang.h

@@ -68,15 +68,14 @@
 #endif
 
 //
-// Driver must call this first, once, before doing any other
-// compiler/linker operations.
+// Call before doing any other compiler/linker operations.
 //
 // (Call once per process, not once per thread.)
 //
 SH_IMPORT_EXPORT int ShInitialize();
 
 //
-// Driver should call this at process shutdown.
+// Call this at process shutdown to clean up memory.
 //
 SH_IMPORT_EXPORT int __fastcall ShFinalize();
 
@@ -290,7 +289,7 @@ SH_IMPORT_EXPORT int ShGetUniformLocation(const ShHandle uniformMap, const char*
 // Deferred-Lowering C++ Interface
 // -----------------------------------
 //
-// Below is a new alternate C++ interface that might potentially replace the above
+// Below is a new alternate C++ interface, which deprecates the above
 // opaque handle-based interface.
 //
 // The below is further designed to handle multiple compilation units per stage, where