more grwl experiments

.gitignore

@@ -4,6 +4,7 @@ tools/mcjitcache
 tools/mcjitcache_release
 tools/publicize
 tools/publicize_release
+
 pyston
 pyston_opt
 pyston_noasserts
@@ -16,6 +17,8 @@ pyston_prof
 pyston_profile
 pyston_release
 pyston_grwl
+pyston_nosync
+
 *.cache
 tests/t.py
 tests/t2.py

src/Makefile

@@ -333,10 +333,10 @@ check:
 	$(MAKE) lint
 	$(MAKE) ext pyston_dbg
 	# $(MAKE) run_unittests
-	$(MAKE) test_dbg
-	$(MAKE) test_release
+	$(MAKE) check_dbg
+	$(MAKE) check_release
 	@# jit_prof forces the use of GCC as the compiler, which can expose other errors, so just build it and see what happens:
-	# $(MAKE) test_prof
+	# $(MAKE) check_prof
 	$(MAKE) pyston_prof
 	$(call checksha,./pyston_prof -cq  $(TESTS_DIR)/raytrace_small.py,0544f4621dd45fe94205219488a2576b84dc044d)
 	$(call checksha,./pyston_prof -cqn $(TESTS_DIR)/raytrace_small.py,0544f4621dd45fe94205219488a2576b84dc044d)
@@ -585,6 +585,7 @@ PASS_OBJS := $(PASS_SRCS:.cpp=.standalone.o)
 $(call make_compile_config,,$(CXXFLAGS_DBG))
 $(call make_compile_config,.release,$(CXXFLAGS_RELEASE))
 $(call make_compile_config,.grwl,$(CXXFLAGS_RELEASE) -DTHREADING_USE_GRWL=1 -DTHREADING_USE_GIL=0 -UBINARY_SUFFIX -DBINARY_SUFFIX=_grwl)
+$(call make_compile_config,.nosync,$(CXXFLAGS_RELEASE) -DTHREADING_USE_GRWL=0 -DTHREADING_USE_GIL=0 -UBINARY_SUFFIX -DBINARY_SUFFIX=_nosync)
 
 $(UNITTEST_SRCS:.cpp=.o): CXXFLAGS += -isystem $(GTEST_DIR)/include
 
@@ -657,6 +658,7 @@ $(call link,_dbg,$(OBJS),$(LDFLAGS),$(LLVM_DEPS))
 $(call link,_debug,$(OBJS),$(LDFLAGS_DEBUG),$(LLVM_DEBUG_DEPS))
 $(call link,_release,$(OPT_OBJS),$(LDFLAGS_RELEASE),$(LLVM_RELEASE_DEPS))
 $(call link,_grwl,stdlib.grwl.bc.o $(SRCS:.cpp=.grwl.o),$(LDFLAGS_RELEASE),$(LLVM_RELEASE_DEPS))
+$(call link,_nosync,stdlib.nosync.bc.o $(SRCS:.cpp=.nosync.o),$(LDFLAGS_RELEASE),$(LLVM_RELEASE_DEPS))
 pyston_oprof: $(OPT_OBJS) codegen/profiling/oprofile.o $(LLVM_DEPS)
 	$(ECHO) Linking $@
 	$(VERB) $(CXX) $(OPT_OBJS) codegen/profiling/oprofile.o $(LDFLAGS_RELEASE) -lopagent -o $@
@@ -695,8 +697,8 @@ RUN_DEPS := ext
 
 define make_target
 $(eval \
-.PHONY: test$1
-test$1: pyston$1 ext
+.PHONY: test$1 check$1
+check$1 test$1: pyston$1 ext
 	python ../tools/tester.py -R pyston$1 -j$(TEST_THREADS) -k $(TESTS_DIR) $(ARGS)
 	python ../tools/tester.py -R pyston$1 -j$(TEST_THREADS) -a -n -k $(TESTS_DIR) $(ARGS)
 	python ../tools/tester.py -R pyston$1 -j$(TEST_THREADS) -a -O -k $(TESTS_DIR) $(ARGS)
@@ -757,6 +759,7 @@ perf_%: perf_release_%
 $(call make_target,_dbg)
 $(call make_target,_release)
 $(call make_target,_grwl)
+$(call make_target,_nosync)
 
 # "kill valgrind":
 kv:

src/core/stats.cpp

@@ -16,6 +16,8 @@
 
 #include <algorithm>
 
+#include "core/thread_utils.h"
+
 namespace pyston {
 
 std::vector<long>* Stats::counts;
@@ -23,6 +25,12 @@ std::unordered_map<int, std::string>* Stats::names;
 StatCounter::StatCounter(const std::string& name) : id(Stats::getStatId(name)) {
 }
 
+StatPerThreadCounter::StatPerThreadCounter(const std::string& name) {
+    char buf[80];
+    snprintf(buf, 80, "%s_t%d", name.c_str(), threading::gettid());
+    id = Stats::getStatId(buf);
+}
+
 int Stats::getStatId(const std::string& name) {
     // hacky but easy way of getting around static constructor ordering issues for now:
     static std::unordered_map<int, std::string> names;

src/core/stats.h

@@ -46,6 +46,17 @@ public:
 
     void log(int count = 1) { Stats::log(id, count); }
 };
+
+struct StatPerThreadCounter {
+private:
+    int id = 0;
+
+public:
+    StatPerThreadCounter(const std::string& name);
+
+    void log(int count = 1) { Stats::log(id, count); }
+};
+
 }
 
 #endif

src/core/thread_utils.h

@@ -18,9 +18,13 @@
 #include <pthread.h>
 #include <unordered_map>
 
+#include "core/common.h"
+
 namespace pyston {
 namespace threading {
 
+pid_t gettid();
+
 template <typename T> class _LockedRegion {
 private:
     T* const mutex;

src/core/threading.cpp

@@ -427,7 +427,7 @@ void allowGLReadPreemption() {
     assert(grwl_state == GRWLHeldState::R);
 
     gl_check_count++;
-    if (gl_check_count < 1000)
+    if (gl_check_count < 10)
         return;
     gl_check_count = 0;
 

src/gc/collector.cpp

@@ -160,14 +160,15 @@ void runCollection() {
     if (VERBOSITY("gc") >= 2)
         printf("Collection #%d\n", ++ncollections);
 
-    threading::GLPromoteRegion _lock;
+    //Timer _t2("promoting", /*min_usec=*/10000);
 
-    Timer _t("collecting", /*min_usec=*/10000);
+    threading::GLPromoteRegion _lock;
 
+    //long promote_us = _t2.end();
+    //static thread_local StatPerThreadCounter sc_promoting_us("gc_promoting_us");
+    //sc_promoting_us.log(promote_us);
 
-    // if (ncollections == 754) {
-    // raise(SIGTRAP);
-    //}
+    Timer _t("collecting", /*min_usec=*/10000);
 
     markPhase();
     sweepPhase();

tools/cumulate.sh

-awk '{if ($0 ~ /%/) { s += $1; print s, "\t", $0 } }'
+awk '{if ($0 ~ /%/) { s += $1; n += $2; _s = $1; $1=""; print s, "\t", _s, " \t", n, "\t", $0 } }'