Purge the old GRWL (gil-free) code

It's time :(

src/codegen/codegen.cpp

@@ -37,8 +37,6 @@
 
 namespace pyston {
 
-DS_DEFINE_RWLOCK(codegen_rwlock);
-
 FunctionMetadata::FunctionMetadata(int num_args, bool takes_varargs, bool takes_kwargs,
                                    std::unique_ptr<SourceInfo> source)
     : code_obj(NULL),

src/codegen/codegen.h

@@ -90,8 +90,6 @@ extern GlobalState g;
 
 // in runtime_hooks.cpp:
 void initGlobalFuncs(GlobalState& g);
-
-DS_DECLARE_RWLOCK(codegen_rwlock);
 }
 
 #endif

src/codegen/irgen/hooks.cpp

@@ -341,9 +341,6 @@ CompiledFunction* compileFunction(FunctionMetadata* f, FunctionSpecialization* s
 void compileAndRunModule(AST_Module* m, BoxedModule* bm) {
     FunctionMetadata* md;
 
-    { // scope for limiting the locked region:
-        LOCK_REGION(codegen_rwlock.asWrite());
-
     Timer _t("for compileModule()");
 
     const char* fn = PyModule_GetFilename(bm);
@@ -364,7 +361,6 @@ void compileAndRunModule(AST_Module* m, BoxedModule* bm) {
         bm->setattr(builtins_str, PyModule_GetDict(builtins_module), NULL);
 
     md = new FunctionMetadata(0, false, false, std::move(si));
-    }
 
     UNAVOIDABLE_STAT_TIMER(t0, "us_timer_interpreted_module_toplevel");
     Box* r = astInterpretFunction(md, NULL, NULL, NULL, NULL, NULL, NULL, NULL);
@@ -395,8 +391,6 @@ Box* evalOrExec(FunctionMetadata* md, Box* globals, Box* boxedLocals) {
 
 static FunctionMetadata* compileForEvalOrExec(AST* source, llvm::ArrayRef<AST_stmt*> body, BoxedString* fn,
                                               PyCompilerFlags* flags) {
-    LOCK_REGION(codegen_rwlock.asWrite());
-
     Timer _t("for evalOrExec()");
 
     ScopingAnalysis* scoping = new ScopingAnalysis(source, false);
@@ -655,8 +649,6 @@ void exec(Box* boxedCode, Box* globals, Box* locals, FutureFlags caller_future_f
 // TODO we should have logic like this at the CLFunc level that detects that we keep
 // on creating functions with failing speculations, and then stop speculating.
 void CompiledFunction::speculationFailed() {
-    LOCK_REGION(codegen_rwlock.asWrite());
-
     this->times_speculation_failed++;
 
     if (this->times_speculation_failed == 4) {
@@ -735,8 +727,6 @@ ConcreteCompilerType* CompiledFunction::getReturnType() {
 /// The cf must be an active version in its parents FunctionMetadata; the given
 /// version will be replaced by the new version, which will be returned.
 static CompiledFunction* _doReopt(CompiledFunction* cf, EffortLevel new_effort) {
-    LOCK_REGION(codegen_rwlock.asWrite());
-
     assert(cf->md->versions.size());
 
     assert(cf);
@@ -773,8 +763,6 @@ static CompiledFunction* _doReopt(CompiledFunction* cf, EffortLevel new_effort)
 static StatCounter stat_osrexits("num_osr_exits");
 static StatCounter stat_osr_compiles("num_osr_compiles");
 CompiledFunction* compilePartialFuncInternal(OSRExit* exit) {
-    LOCK_REGION(codegen_rwlock.asWrite());
-
     assert(exit);
     stat_osrexits.log();
 

src/core/threading.cpp

@@ -36,6 +36,8 @@
 namespace pyston {
 namespace threading {
 
+void _acquireGIL();
+void _releaseGIL();
 
 #ifdef WITH_THREAD
 #include "pythread.h"
@@ -99,14 +101,14 @@ public:
         return holds_gil;
     }
 
-    void takeGil() {
+    void gilTaken() {
         assert(pthread_self() == this->pthread_id);
 
         assert(!holds_gil);
         holds_gil = true;
     }
 
-    void releaseGil()  {
+    void gilReleased() {
         assert(pthread_self() == this->pthread_id);
 
         assert(holds_gil);
@@ -197,7 +199,7 @@ extern "C" PyGILState_STATE PyGILState_Ensure(void) noexcept {
         if (current_internal_thread_state == NULL)
             Py_FatalError("Couldn't create thread-state for new thread");
 
-        acquireGLRead();
+        _acquireGIL();
         return PyGILState_UNLOCKED;
     } else {
         ++cur_thread_state.gilstate_counter;
@@ -246,13 +248,14 @@ static void* _thread_start(void* _arg) {
     Box* arg3 = arg->arg3;
     delete arg;
 
-    threading::GLReadRegion _glock;
+    _acquireGIL();
     registerThread(true);
     assert(!PyErr_Occurred());
 
     void* rtn = start_func(arg1, arg2, arg3);
 
     unregisterThread();
+    _releaseGIL();
 
     return rtn;
 }
@@ -296,6 +299,8 @@ void registerMainThread() {
     assert(!current_internal_thread_state);
     current_internal_thread_state = new ThreadStateInternal(pthread_self(), &cur_thread_state);
     current_threads[pthread_self()] = current_internal_thread_state;
+
+    _acquireGIL();
 }
 
 /* Wait until threading._shutdown completes, provided
@@ -344,28 +349,23 @@ extern "C" void beginAllowThreads() noexcept {
         LOCK_REGION(&threading_lock);
 
         assert(current_internal_thread_state);
-        current_internal_thread_state->releaseGil();
+        current_internal_thread_state->gilReleased();
     }
 
-    releaseGLRead();
+    _releaseGIL();
 }
 
 extern "C" void endAllowThreads() noexcept {
-    acquireGLRead();
+    _acquireGIL();
 
     {
         LOCK_REGION(&threading_lock);
 
         assert(current_internal_thread_state);
-        current_internal_thread_state->takeGil();
+        current_internal_thread_state->gilTaken();
     }
 }
 
-#if THREADING_USE_GIL
-#if THREADING_USE_GRWL
-#error "Can't turn on both the GIL and the GRWL!"
-#endif
-
 static pthread_mutex_t gil = PTHREAD_MUTEX_INITIALIZER;
 
 std::atomic<int> threads_waiting_on_gil(0);
@@ -423,7 +423,7 @@ extern "C" void PyEval_ReInitThreads() noexcept {
     Py_DECREF(threading);
 }
 
-void acquireGLWrite() {
+void _acquireGIL() {
     threads_waiting_on_gil++;
     pthread_mutex_lock(&gil);
     threads_waiting_on_gil--;
@@ -431,7 +431,7 @@ void acquireGLWrite() {
     pthread_cond_signal(&gil_acquired);
 }
 
-void releaseGLWrite() {
+void _releaseGIL() {
     pthread_mutex_unlock(&gil);
 }
 
@@ -457,86 +457,6 @@ void _allowGLReadPreemption() {
     threads_waiting_on_gil--;
     pthread_cond_signal(&gil_acquired);
 }
-#elif THREADING_USE_GRWL
-static pthread_rwlock_t grwl = PTHREAD_RWLOCK_WRITER_NONRECURSIVE_INITIALIZER_NP;
-
-enum class GRWLHeldState {
-    N,
-    R,
-    W,
-};
-static __thread GRWLHeldState grwl_state = GRWLHeldState::N;
-
-static std::atomic<int> writers_waiting(0);
-
-void acquireGLRead() {
-    assert(grwl_state == GRWLHeldState::N);
-    pthread_rwlock_rdlock(&grwl);
-    grwl_state = GRWLHeldState::R;
-}
-
-void releaseGLRead() {
-    assert(grwl_state == GRWLHeldState::R);
-    pthread_rwlock_unlock(&grwl);
-    grwl_state = GRWLHeldState::N;
-}
-
-void acquireGLWrite() {
-    assert(grwl_state == GRWLHeldState::N);
-
-    writers_waiting++;
-    pthread_rwlock_wrlock(&grwl);
-    writers_waiting--;
-
-    grwl_state = GRWLHeldState::W;
-}
-
-void releaseGLWrite() {
-    assert(grwl_state == GRWLHeldState::W);
-    pthread_rwlock_unlock(&grwl);
-    grwl_state = GRWLHeldState::N;
-}
-
-void promoteGL() {
-    Timer _t2("promoting", /*min_usec=*/10000);
-
-    // Note: this is *not* the same semantics as normal promoting, on purpose.
-    releaseGLRead();
-    acquireGLWrite();
-
-    long promote_us = _t2.end();
-    static thread_local StatPerThreadCounter sc_promoting_us("grwl_promoting_us");
-    sc_promoting_us.log(promote_us);
-}
-
-void demoteGL() {
-    releaseGLWrite();
-    acquireGLRead();
-}
-
-static __thread int gl_check_count = 0;
-void allowGLReadPreemption() {
-    assert(grwl_state == GRWLHeldState::R);
-
-    // gl_check_count++;
-    // if (gl_check_count < 10)
-    // return;
-    // gl_check_count = 0;
-
-    if (__builtin_expect(!writers_waiting.load(std::memory_order_relaxed), 1))
-        return;
-
-    Timer _t2("preempted", /*min_usec=*/10000);
-    pthread_rwlock_unlock(&grwl);
-    // The GRWL is a writer-prefered rwlock, so this next statement will block even
-    // if the lock is in read mode:
-    pthread_rwlock_rdlock(&grwl);
-
-    long preempt_us = _t2.end();
-    static thread_local StatPerThreadCounter sc_preempting_us("grwl_preempt_us");
-    sc_preempting_us.log(preempt_us);
-}
-#endif
 
 // We don't support CPython's TLS (yet?)
 extern "C" void PyThread_ReInitTLS(void) noexcept {

src/core/threading.h

@@ -47,35 +47,10 @@ void finishMainThread();
 
 bool isMainThread();
 
-#ifndef THREADING_USE_GIL
-#define THREADING_USE_GIL 1
-#define THREADING_USE_GRWL 0
-#endif
-#define THREADING_SAFE_DATASTRUCTURES THREADING_USE_GRWL
-
-#if THREADING_SAFE_DATASTRUCTURES
-#define DS_DEFINE_MUTEX(name) pyston::threading::PthreadFastMutex name
-
-#define DS_DECLARE_RWLOCK(name) extern pyston::threading::PthreadRWLock name
-#define DS_DEFINE_RWLOCK(name) pyston::threading::PthreadRWLock name
-
-#define DS_DEFINE_SPINLOCK(name) pyston::threading::PthreadSpinLock name
-#else
-#define DS_DEFINE_MUTEX(name) pyston::threading::NopLock name
-
-#define DS_DECLARE_RWLOCK(name) extern pyston::threading::NopLock name
-#define DS_DEFINE_RWLOCK(name) pyston::threading::NopLock name
-
-#define DS_DEFINE_SPINLOCK(name) pyston::threading::NopLock name
-#endif
-
-void acquireGLRead();
-void releaseGLRead();
-void acquireGLWrite();
-void releaseGLWrite();
 void _allowGLReadPreemption();
 
 #define GIL_CHECK_INTERVAL 1000
+
 // Note: this doesn't need to be an atomic, since it should
 // only be accessed by the thread that holds the gil:
 extern int gil_check_count;
@@ -104,25 +79,7 @@ extern "C" inline void allowGLReadPreemption() {
 
     _allowGLReadPreemption();
 }
-// Note: promoteGL is free to drop the lock and then reacquire
-void promoteGL();
-void demoteGL();
-
-
-
-// Helper macro for creating a RAII wrapper around two functions.
-#define MAKE_REGION(name, start, end)                                                                                  \
-    class name {                                                                                                       \
-    public:                                                                                                            \
-        name() { start(); }                                                                                            \
-        ~name() { end(); }                                                                                             \
-    };
 
-MAKE_REGION(GLReadRegion, acquireGLRead, releaseGLRead);
-MAKE_REGION(GLPromoteRegion, promoteGL, demoteGL);
-// MAKE_REGION(GLReadReleaseRegion, releaseGLRead, acquireGLRead);
-// MAKE_REGION(GLWriteReleaseRegion, releaseGLWrite, acquireGLWrite);
-#undef MAKE_REGION
 
 extern "C" void beginAllowThreads() noexcept;
 extern "C" void endAllowThreads() noexcept;
@@ -134,37 +91,6 @@ public:
 };
 
 
-#if THREADING_USE_GIL
-inline void acquireGLRead() {
-    acquireGLWrite();
-}
-inline void releaseGLRead() {
-    releaseGLWrite();
-}
-inline void promoteGL() {
-}
-inline void demoteGL() {
-}
-#endif
-
-#if !THREADING_USE_GIL && !THREADING_USE_GRWL
-inline void acquireGLRead() {
-}
-inline void releaseGLRead() {
-}
-inline void acquireGLWrite() {
-}
-inline void releaseGLWrite() {
-}
-inline void promoteGL() {
-}
-inline void demoteGL() {
-}
-extern "C" inline void allowGLReadPreemption() __attribute__((visibility("default")));
-extern "C" inline void allowGLReadPreemption() {
-}
-#endif
-
 extern bool forgot_refs_via_fork;
 
 } // namespace threading

src/jit.cpp

@@ -369,7 +369,6 @@ static int main(int argc, char** argv) noexcept {
         const char* fn = NULL;
 
         threading::registerMainThread();
-        threading::acquireGLRead();
 
         Py_SetProgramName(argv[0]);
 
@@ -540,11 +539,6 @@ static int main(int argc, char** argv) noexcept {
 
         threading::finishMainThread();
 
-        // Acquire the GIL to make sure we stop the other threads, since we will tear down
-        // data structures they are potentially running on.
-        // Note: we will purposefully not release the GIL on exiting.
-        threading::promoteGL();
-
         _t.split("Py_Finalize");
 
         Py_Finalize();

src/runtime/objmodel.cpp

@@ -3967,8 +3967,6 @@ static StatCounter slowpath_pickversion("slowpath_pickversion");
 template <ExceptionStyle S>
 static CompiledFunction* pickVersion(FunctionMetadata* f, int num_output_args, Box* oarg1, Box* oarg2, Box* oarg3,
                                      Box** oargs) {
-    LOCK_REGION(codegen_rwlock.asWrite());
-
     // if always_use_version is set use it even if the exception style does not match.
     // But prefer using the correct style if both are available
     if (f->always_use_version.get(S))

test/unittests/unittests.h

@@ -25,12 +25,10 @@ namespace pyston {
 class PystonTestEnvironment : public testing::Environment {
     void SetUp() override {
         threading::registerMainThread();
-
-        threading::acquireGLRead();
     }
 
     void TearDown() override {
-        threading::releaseGLRead();
+        // threading::releaseGLRead();
     }
 };