vendor/compiler-rt/compiler-rt-r182741

113 files changed, 3738 insertions, 1204 deletions

diff --git a/lib/tsan/CMakeLists.txt b/lib/tsan/CMakeLists.txt
index 34e3a2ea524e2..2828895675099 100644
--- a/lib/tsan/CMakeLists.txt
+++ b/lib/tsan/CMakeLists.txt
@@ -2,7 +2,12 @@
 
 include_directories(..)
 
-set(TSAN_CFLAGS ${SANITIZER_COMMON_CFLAGS})
+# SANITIZER_COMMON_CFLAGS contains -fPIC, but it's performance-critical for
+# TSan runtime to be built with -fPIE to reduce the number of register spills.
+set(TSAN_CFLAGS
+  ${SANITIZER_COMMON_CFLAGS}
+  -fPIE
+  -fno-rtti)
 # FIXME: Add support for compile flags:
 #   -Wframe-larger-than=512,
 #   -Wglobal-constructors,
diff --git a/lib/tsan/Makefile.old b/lib/tsan/Makefile.old
index 593482fbb5da4..b548f5d2f6eed 100644
--- a/lib/tsan/Makefile.old
+++ b/lib/tsan/Makefile.old
@@ -1,13 +1,16 @@
 DEBUG=0
 LDFLAGS=-ldl -lpthread -pie
-CXXFLAGS = -fPIE -g -Wall -Werror -DTSAN_DEBUG=$(DEBUG) -DSANITIZER_DEBUG=$(DEBUG)
+CXXFLAGS = -fPIE -fno-rtti -g -Wall -Werror \
+					 -DGTEST_HAS_RTTI=0 -DTSAN_DEBUG=$(DEBUG) -DSANITIZER_DEBUG=$(DEBUG)
+CLANG=clang
+FILECHECK=FileCheck
 # Silence warnings that Clang produces for gtest code.
 # Use -Wno-attributes so that gcc doesn't complain about unknown warning types.
 CXXFLAGS += -Wno-attributes
 ifeq ($(DEBUG), 0)
 	CXXFLAGS += -O3
 endif
-ifeq ($(CXX), clang++)
+ifeq ($(CXX), $(CLANG)++)
   CXXFLAGS+= -Wno-unused-private-field -Wno-static-in-inline -Wgnu
 endif
 
@@ -54,16 +57,16 @@ test: libtsan tsan_test
 
 run: all
 	(ulimit -s 8192; ./tsan_test)
-	./lit_tests/test_output.sh
+	CC=$(CLANG) CXX=$(CLANG)++ FILECHECK=$(FILECHECK) ./lit_tests/test_output.sh
 
 presubmit:
 	../sanitizer_common/scripts/check_lint.sh
 	# Debug build with clang.
 	$(MAKE) -f Makefile.old clean
-	$(MAKE) -f Makefile.old run DEBUG=1 -j 16 CC=clang CXX=clang++
+	$(MAKE) -f Makefile.old run DEBUG=1 -j 16 CC=$(CLANG) CXX=$(CLANG)++
 	# Release build with clang.
 	$(MAKE) -f Makefile.old clean
-	$(MAKE) -f Makefile.old run DEBUG=0 -j 16 CC=clang CXX=clang++
+	$(MAKE) -f Makefile.old run DEBUG=0 -j 16 CC=$(CLANG) CXX=$(CLANG)++
 	# Debug build with gcc
 	$(MAKE) -f Makefile.old clean
 	$(MAKE) -f Makefile.old run DEBUG=1 -j 16 CC=gcc CXX=g++
@@ -93,3 +96,5 @@ clean:
 	rm -f asm_*.s libtsan.nm libtsan.objdump */*.o tsan_test
 	rm -rf $(GTEST_BUILD_DIR)
 	$(MAKE) clean -C rtl -f Makefile.old
+	rm -f go/*.s
+	rm -rf build
diff --git a/lib/tsan/analyze_libtsan.sh b/lib/tsan/analyze_libtsan.sh
index e0805610714b1..705e4c5460f2e 100755
--- a/lib/tsan/analyze_libtsan.sh
+++ b/lib/tsan/analyze_libtsan.sh
@@ -4,7 +4,7 @@ set -e
 set -u
 
 get_asm() {
-  grep tsan_$1.: -A 10000 libtsan.objdump | \
+  grep __tsan_$1.: -A 10000 libtsan.objdump | \
     awk "/[^:]$/ {print;} />:/ {c++; if (c == 2) {exit}}"
 }
 
@@ -27,7 +27,7 @@ for f in $list; do
   file=asm_$f.s
   get_asm $f > $file
   tot=$(wc -l < $file)
-  size=$(grep $f$ libtsan.nm | awk --non-decimal-data '{print ("0x"$2)+0}')
+  size=$(grep __tsan_$f$ libtsan.nm | awk --non-decimal-data '{print ("0x"$2)+0}')
   rsp=$(grep '(%rsp)' $file | wc -l)
   push=$(grep 'push' $file | wc -l)
   pop=$(grep 'pop' $file | wc -l)
diff --git a/lib/tsan/check_cmake.sh b/lib/tsan/check_cmake.sh
index 5f11e727f0916..52c97c339096b 100755
--- a/lib/tsan/check_cmake.sh
+++ b/lib/tsan/check_cmake.sh
@@ -7,5 +7,6 @@ mkdir -p $ROOT/build
 cd $ROOT/build
 CC=clang CXX=clang++ cmake -DLLVM_ENABLE_WERROR=ON -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON $ROOT/../../../..
 make -j64
-make check-tsan check-sanitizer -j64
-
+make check-sanitizer -j64
+make check-tsan -j64
+make check-asan -j64
diff --git a/lib/tsan/go/buildgo.sh b/lib/tsan/go/buildgo.sh
index a153afd6ee8eb..51f1a7975b573 100755
--- a/lib/tsan/go/buildgo.sh
+++ b/lib/tsan/go/buildgo.sh
@@ -20,6 +20,7 @@ SRCS="
 	../../sanitizer_common/sanitizer_flags.cc
 	../../sanitizer_common/sanitizer_libc.cc
 	../../sanitizer_common/sanitizer_printf.cc
+	../../sanitizer_common/sanitizer_thread_registry.cc
 "
 
 if [ "`uname -a | grep Linux`" != "" ]; then
@@ -29,7 +30,9 @@ if [ "`uname -a | grep Linux`" != "" ]; then
 	SRCS+="
 		../rtl/tsan_platform_linux.cc
 		../../sanitizer_common/sanitizer_posix.cc
+		../../sanitizer_common/sanitizer_posix_libcdep.cc
 		../../sanitizer_common/sanitizer_linux.cc
+		../../sanitizer_common/sanitizer_linux_libcdep.cc
 	"
 elif [ "`uname -a | grep Darwin`" != "" ]; then
 	SUFFIX="darwin_amd64"
@@ -60,7 +63,7 @@ for F in $SRCS; do
 	cat $F >> gotsan.cc
 done
 
-FLAGS=" -I../rtl -I../.. -I../../sanitizer_common -I../../../include -m64 -Wall -Werror -fno-exceptions -DTSAN_GO -DSANITIZER_GO -DTSAN_SHADOW_COUNT=4 $OSCFLAGS"
+FLAGS=" -I../rtl -I../.. -I../../sanitizer_common -I../../../include -m64 -Wall -Werror -fno-exceptions -fno-rtti -DTSAN_GO -DSANITIZER_GO -DTSAN_SHADOW_COUNT=4 $OSCFLAGS"
 if [ "$DEBUG" == "" ]; then
 	FLAGS+=" -DTSAN_DEBUG=0 -O3 -fomit-frame-pointer"
 else
diff --git a/lib/tsan/go/test.c b/lib/tsan/go/test.c
index 2414a1e9925f7..902dfc915582b 100644
--- a/lib/tsan/go/test.c
+++ b/lib/tsan/go/test.c
@@ -13,20 +13,20 @@
 
 #include <stdio.h>
 
-void __tsan_init();
+void __tsan_init(void **thr);
 void __tsan_fini();
 void __tsan_map_shadow(void *addr, unsigned long size);
-void __tsan_go_start(int pgoid, int chgoid, void *pc);
-void __tsan_go_end(int goid);
-void __tsan_read(int goid, void *addr, void *pc);
-void __tsan_write(int goid, void *addr, void *pc);
-void __tsan_func_enter(int goid, void *pc);
-void __tsan_func_exit(int goid);
-void __tsan_malloc(int goid, void *p, unsigned long sz, void *pc);
+void __tsan_go_start(void *thr, void **chthr, void *pc);
+void __tsan_go_end(void *thr);
+void __tsan_read(void *thr, void *addr, void *pc);
+void __tsan_write(void *thr, void *addr, void *pc);
+void __tsan_func_enter(void *thr, void *pc);
+void __tsan_func_exit(void *thr);
+void __tsan_malloc(void *thr, void *p, unsigned long sz, void *pc);
 void __tsan_free(void *p);
-void __tsan_acquire(int goid, void *addr);
-void __tsan_release(int goid, void *addr);
-void __tsan_release_merge(int goid, void *addr);
+void __tsan_acquire(void *thr, void *addr);
+void __tsan_release(void *thr, void *addr);
+void __tsan_release_merge(void *thr, void *addr);
 
 int __tsan_symbolize(void *pc, char **img, char **rtn, char **file, int *l) {
   return 0;
@@ -35,19 +35,21 @@ int __tsan_symbolize(void *pc, char **img, char **rtn, char **file, int *l) {
 char buf[10];
 
 int main(void) {
-  __tsan_init();
+  void *thr0 = 0;
+  __tsan_init(&thr0);
   __tsan_map_shadow(buf, sizeof(buf) + 4096);
-  __tsan_func_enter(0, &main);
-  __tsan_malloc(0, buf, 10, 0);
-  __tsan_release(0, buf);
-  __tsan_release_merge(0, buf);
-  __tsan_go_start(0, 1, 0);
-  __tsan_write(1, buf, 0);
-  __tsan_acquire(1, buf);
-  __tsan_go_end(1);
-  __tsan_read(0, buf, 0);
+  __tsan_func_enter(thr0, &main);
+  __tsan_malloc(thr0, buf, 10, 0);
+  __tsan_release(thr0, buf);
+  __tsan_release_merge(thr0, buf);
+  void *thr1 = 0;
+  __tsan_go_start(thr0, &thr1, 0);
+  __tsan_write(thr1, buf, 0);
+  __tsan_acquire(thr1, buf);
+  __tsan_go_end(thr1);
+  __tsan_read(thr0, buf, 0);
   __tsan_free(buf);
-  __tsan_func_exit(0);
+  __tsan_func_exit(thr0);
   __tsan_fini();
   return 0;
 }
diff --git a/lib/tsan/go/tsan_go.cc b/lib/tsan/go/tsan_go.cc
index 360608a0cf1ba..957d58211281e 100644
--- a/lib/tsan/go/tsan_go.cc
+++ b/lib/tsan/go/tsan_go.cc
@@ -18,10 +18,6 @@
 
 namespace __tsan {
 
-const int kMaxGoroutinesEver = 128*1024;
-
-static ThreadState *goroutines[kMaxGoroutinesEver];
-
 void InitializeInterceptors() {
 }
 
@@ -80,20 +76,18 @@ ReportStack *SymbolizeCode(uptr addr) {
 
 extern "C" {
 
-static void AllocGoroutine(int tid) {
-  if (tid >= kMaxGoroutinesEver) {
-    Printf("FATAL: Reached goroutine limit\n");
-    Die();
-  }
+static ThreadState *main_thr;
+
+static ThreadState *AllocGoroutine() {
   ThreadState *thr = (ThreadState*)internal_alloc(MBlockThreadContex,
       sizeof(ThreadState));
   internal_memset(thr, 0, sizeof(*thr));
-  goroutines[tid] = thr;
+  return thr;
 }
 
-void __tsan_init() {
-  AllocGoroutine(0);
-  ThreadState *thr = goroutines[0];
+void __tsan_init(ThreadState **thrp) {
+  ThreadState *thr = AllocGoroutine();
+  main_thr = *thrp = thr;
   thr->in_rtl++;
   Initialize(thr);
   thr->in_rtl--;
@@ -101,7 +95,7 @@ void __tsan_init() {
 
 void __tsan_fini() {
   // FIXME: Not necessary thread 0.
-  ThreadState *thr = goroutines[0];
+  ThreadState *thr = main_thr;
   thr->in_rtl++;
   int res = Finalize(thr);
   thr->in_rtl--;
@@ -112,44 +106,37 @@ void __tsan_map_shadow(uptr addr, uptr size) {
   MapShadow(addr, size);
 }
 
-void __tsan_read(int goid, void *addr, void *pc) {
-  ThreadState *thr = goroutines[goid];
-  MemoryAccess(thr, (uptr)pc, (uptr)addr, 0, false);
+void __tsan_read(ThreadState *thr, void *addr, void *pc) {
+  MemoryRead(thr, (uptr)pc, (uptr)addr, kSizeLog1);
 }
 
-void __tsan_write(int goid, void *addr, void *pc) {
-  ThreadState *thr = goroutines[goid];
-  MemoryAccess(thr, (uptr)pc, (uptr)addr, 0, true);
+void __tsan_write(ThreadState *thr, void *addr, void *pc) {
+  MemoryWrite(thr, (uptr)pc, (uptr)addr, kSizeLog1);
 }
 
-void __tsan_read_range(int goid, void *addr, uptr size, uptr step, void *pc) {
-  ThreadState *thr = goroutines[goid];
-  for (uptr i = 0; i < size; i += step)
-	  MemoryAccess(thr, (uptr)pc, (uptr)addr + i, 0, false);
+void __tsan_read_range(ThreadState *thr, void *addr, uptr size, uptr step,
+                       void *pc) {
+  MemoryAccessRangeStep(thr, (uptr)pc, (uptr)addr, size, step, false);
 }
 
-void __tsan_write_range(int goid, void *addr, uptr size, uptr step, void *pc) {
-  ThreadState *thr = goroutines[goid];
-  for (uptr i = 0; i < size; i += step)
-	  MemoryAccess(thr, (uptr)pc, (uptr)addr + i, 0, true);
+void __tsan_write_range(ThreadState *thr, void *addr, uptr size, uptr step,
+                        void *pc) {
+  MemoryAccessRangeStep(thr, (uptr)pc, (uptr)addr, size, step, true);
 }
 
-void __tsan_func_enter(int goid, void *pc) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_func_enter(ThreadState *thr, void *pc) {
   FuncEntry(thr, (uptr)pc);
 }
 
-void __tsan_func_exit(int goid) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_func_exit(ThreadState *thr) {
   FuncExit(thr);
 }
 
-void __tsan_malloc(int goid, void *p, uptr sz, void *pc) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_malloc(ThreadState *thr, void *p, uptr sz, void *pc) {
   if (thr == 0)  // probably before __tsan_init()
     return;
   thr->in_rtl++;
-  MemoryRangeImitateWrite(thr, (uptr)pc, (uptr)p, sz);
+  MemoryResetRange(thr, (uptr)pc, (uptr)p, sz);
   thr->in_rtl--;
 }
 
@@ -157,56 +144,47 @@ void __tsan_free(void *p) {
   (void)p;
 }
 
-void __tsan_go_start(int pgoid, int chgoid, void *pc) {
-  if (chgoid == 0)
-    return;
-  AllocGoroutine(chgoid);
-  ThreadState *thr = goroutines[chgoid];
-  ThreadState *parent = goroutines[pgoid];
+void __tsan_go_start(ThreadState *parent, ThreadState **pthr, void *pc) {
+  ThreadState *thr = AllocGoroutine();
+  *pthr = thr;
   thr->in_rtl++;
   parent->in_rtl++;
-  int goid2 = ThreadCreate(parent, (uptr)pc, 0, true);
-  ThreadStart(thr, goid2, 0);
+  int goid = ThreadCreate(parent, (uptr)pc, 0, true);
+  ThreadStart(thr, goid, 0);
   parent->in_rtl--;
   thr->in_rtl--;
 }
 
-void __tsan_go_end(int goid) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_go_end(ThreadState *thr) {
   thr->in_rtl++;
   ThreadFinish(thr);
   thr->in_rtl--;
   internal_free(thr);
-  goroutines[goid] = 0;
 }
 
-void __tsan_acquire(int goid, void *addr) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_acquire(ThreadState *thr, void *addr) {
   thr->in_rtl++;
   Acquire(thr, 0, (uptr)addr);
   thr->in_rtl--;
 }
 
-void __tsan_release(int goid, void *addr) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_release(ThreadState *thr, void *addr) {
   thr->in_rtl++;
   ReleaseStore(thr, 0, (uptr)addr);
   thr->in_rtl--;
 }
 
-void __tsan_release_merge(int goid, void *addr) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_release_merge(ThreadState *thr, void *addr) {
   thr->in_rtl++;
   Release(thr, 0, (uptr)addr);
   thr->in_rtl--;
 }
 
-void __tsan_finalizer_goroutine(int goid) {
-  ThreadState *thr = goroutines[goid];
+void __tsan_finalizer_goroutine(ThreadState *thr) {
   AcquireGlobal(thr, 0);
 }
 
-#ifdef _WIN32
+#if SANITIZER_WINDOWS
 // MinGW gcc emits calls to the function.
 void ___chkstk_ms(void) {
 // The implementation must be along the lines of:
@@ -242,3 +220,11 @@ void ___chkstk_ms(void) {
 
 }  // extern "C"
 }  // namespace __tsan
+
+namespace __sanitizer {
+
+void SymbolizerPrepareForSandboxing() {
+  // Nothing to do here for Go.
+}
+
+}  // namespace __sanitizer
diff --git a/lib/tsan/lit_tests/CMakeLists.txt b/lib/tsan/lit_tests/CMakeLists.txt
index ff2508dd75afd..53e5015d1bc4d 100644
--- a/lib/tsan/lit_tests/CMakeLists.txt
+++ b/lib/tsan/lit_tests/CMakeLists.txt
@@ -11,9 +11,8 @@ configure_lit_site_cfg(
 if(COMPILER_RT_CAN_EXECUTE_TESTS)
   # Run TSan output tests only if we're sure we can produce working binaries.
   set(TSAN_TEST_DEPS
-    clang clang-headers FileCheck count not llvm-symbolizer
-    ${TSAN_RUNTIME_LIBRARIES}
-    )
+    ${SANITIZER_COMMON_LIT_TEST_DEPS}
+    ${TSAN_RUNTIME_LIBRARIES})
   set(TSAN_TEST_PARAMS
     tsan_site_config=${CMAKE_CURRENT_BINARY_DIR}/lit.site.cfg
     )
diff --git a/lib/tsan/lit_tests/SharedLibs/lit.local.cfg b/lib/tsan/lit_tests/SharedLibs/lit.local.cfg
new file mode 100644
index 0000000000000..b3677c17a0f2a
--- /dev/null
+++ b/lib/tsan/lit_tests/SharedLibs/lit.local.cfg
@@ -0,0 +1,4 @@
+# Sources in this directory are compiled as shared libraries and used by
+# tests in parent directory.
+
+config.suffixes = []
diff --git a/lib/tsan/lit_tests/SharedLibs/load_shared_lib-so.cc b/lib/tsan/lit_tests/SharedLibs/load_shared_lib-so.cc
new file mode 100644
index 0000000000000..d05aa6a40d182
--- /dev/null
+++ b/lib/tsan/lit_tests/SharedLibs/load_shared_lib-so.cc
@@ -0,0 +1,22 @@
+//===----------- load_shared_lib-so.cc --------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file is a part of ThreadSanitizer (TSan), a race detector.
+//
+//===----------------------------------------------------------------------===//
+
+#include <stddef.h>
+
+int GLOB_SHARED = 0;
+
+extern "C"
+void *write_from_so(void *unused) {
+  GLOB_SHARED++;
+  return NULL;
+}
diff --git a/lib/tsan/lit_tests/Unit/lit.cfg b/lib/tsan/lit_tests/Unit/lit.cfg
index 6688697c0c1b1..0a0dbbfa54950 100644
--- a/lib/tsan/lit_tests/Unit/lit.cfg
+++ b/lib/tsan/lit_tests/Unit/lit.cfg
@@ -11,9 +11,8 @@ def get_required_attr(config, attr_name):
   return attr_value
 
 # Setup attributes common for all compiler-rt projects.
-llvm_src_root = get_required_attr(config, 'llvm_src_root')
-compiler_rt_lit_unit_cfg = os.path.join(llvm_src_root, "projects",
-                                        "compiler-rt", "lib",
+compiler_rt_src_root = get_required_attr(config, 'compiler_rt_src_root')
+compiler_rt_lit_unit_cfg = os.path.join(compiler_rt_src_root, "lib",
                                         "lit.common.unit.cfg")
 lit.load_config(config, compiler_rt_lit_unit_cfg)
 
diff --git a/lib/tsan/lit_tests/Unit/lit.site.cfg.in b/lib/tsan/lit_tests/Unit/lit.site.cfg.in
index 23654b9be2eee..6eedc21808765 100644
--- a/lib/tsan/lit_tests/Unit/lit.site.cfg.in
+++ b/lib/tsan/lit_tests/Unit/lit.site.cfg.in
@@ -1,15 +1,17 @@
 ## Autogenerated by LLVM/Clang configuration.
 # Do not edit!
 
-config.build_type = "@CMAKE_BUILD_TYPE@"
 config.llvm_obj_root = "@LLVM_BINARY_DIR@"
 config.llvm_src_root = "@LLVM_SOURCE_DIR@"
+config.compiler_rt_src_root = "@COMPILER_RT_SOURCE_DIR@"
 config.llvm_tools_dir = "@LLVM_TOOLS_DIR@"
+config.llvm_build_mode = "@LLVM_BUILD_MODE@"
 
 # LLVM tools dir can be passed in lit parameters, so try to
 # apply substitution.
 try:
   config.llvm_tools_dir = config.llvm_tools_dir % lit.params
+  config.llvm_build_mode = config.llvm_build_mode % lit.params
 except KeyError,e:
   key, = e.args
   lit.fatal("unable to find %r parameter, use '--param=%s=VALUE'" % (key, key))
diff --git a/lib/tsan/lit_tests/aligned_vs_unaligned_race.cc b/lib/tsan/lit_tests/aligned_vs_unaligned_race.cc
new file mode 100644
index 0000000000000..f4533d08306c5
--- /dev/null
+++ b/lib/tsan/lit_tests/aligned_vs_unaligned_race.cc
@@ -0,0 +1,34 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+// Race between an aligned access and an unaligned access, which
+// touches the same memory region.
+// This is a real race which is not detected by tsan.
+// https://code.google.com/p/thread-sanitizer/issues/detail?id=17
+#include <pthread.h>
+#include <stdio.h>
+#include <stdint.h>
+
+uint64_t Global[2];
+
+void *Thread1(void *x) {
+  Global[1]++;
+  return NULL;
+}
+
+void *Thread2(void *x) {
+  char *p1 = reinterpret_cast<char *>(&Global[0]);
+  uint64_t *p4 = reinterpret_cast<uint64_t *>(p1 + 1);
+  (*p4)++;
+  return NULL;
+}
+
+int main() {
+  pthread_t t[2];
+  pthread_create(&t[0], NULL, Thread1, NULL);
+  pthread_create(&t[1], NULL, Thread2, NULL);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+  printf("Pass\n");
+  // CHECK-NOT: ThreadSanitizer: data race
+  // CHECK: Pass
+  return 0;
+}
diff --git a/lib/tsan/lit_tests/atomic_free.cc b/lib/tsan/lit_tests/atomic_free.cc
new file mode 100644
index 0000000000000..ba9bd5ac4aedd
--- /dev/null
+++ b/lib/tsan/lit_tests/atomic_free.cc
@@ -0,0 +1,19 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <unistd.h>
+
+void *Thread(void *a) {
+  __atomic_fetch_add((int*)a, 1, __ATOMIC_SEQ_CST);
+  return 0;
+}
+
+int main() {
+  int *a = new int(0);
+  pthread_t t;
+  pthread_create(&t, 0, Thread, a);
+  sleep(1);
+  delete a;
+  pthread_join(t, 0);
+}
+
+// CHECK: WARNING: ThreadSanitizer: data race
diff --git a/lib/tsan/lit_tests/atomic_free2.cc b/lib/tsan/lit_tests/atomic_free2.cc
new file mode 100644
index 0000000000000..5517bf7ce9029
--- /dev/null
+++ b/lib/tsan/lit_tests/atomic_free2.cc
@@ -0,0 +1,19 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <unistd.h>
+
+void *Thread(void *a) {
+  sleep(1);
+  __atomic_fetch_add((int*)a, 1, __ATOMIC_SEQ_CST);
+  return 0;
+}
+
+int main() {
+  int *a = new int(0);
+  pthread_t t;
+  pthread_create(&t, 0, Thread, a);
+  delete a;
+  pthread_join(t, 0);
+}
+
+// CHECK: WARNING: ThreadSanitizer: heap-use-after-free
diff --git a/lib/tsan/lit_tests/atomic_norace.cc b/lib/tsan/lit_tests/atomic_norace.cc
new file mode 100644
index 0000000000000..265459b0758e1
--- /dev/null
+++ b/lib/tsan/lit_tests/atomic_norace.cc
@@ -0,0 +1,61 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <unistd.h>
+
+const int kTestCount = 4;
+typedef long long T;
+T atomics[kTestCount * 2];
+
+void Test(int test, T *p, bool main_thread) {
+  volatile T sink;
+  if (test == 0) {
+    if (main_thread)
+      __atomic_fetch_add(p, 1, __ATOMIC_RELAXED);
+    else
+      __atomic_fetch_add(p, 1, __ATOMIC_RELAXED);
+  } else if (test == 1) {
+    if (main_thread)
+      __atomic_exchange_n(p, 1, __ATOMIC_ACQ_REL);
+    else
+      __atomic_exchange_n(p, 1, __ATOMIC_ACQ_REL);
+  } else if (test == 2) {
+    if (main_thread)
+      sink = __atomic_load_n(p, __ATOMIC_SEQ_CST);
+    else
+      __atomic_store_n(p, 1, __ATOMIC_SEQ_CST);
+  } else if (test == 3) {
+    if (main_thread)
+      sink = __atomic_load_n(p, __ATOMIC_SEQ_CST);
+    else
+      sink = *p;
+  }
+}
+
+void *Thread(void *p) {
+  for (int i = 0; i < kTestCount; i++) {
+    Test(i, &atomics[i], false);
+  }
+  sleep(2);
+  for (int i = 0; i < kTestCount; i++) {
+    fprintf(stderr, "Test %d reverse\n", i);
+    Test(i, &atomics[kTestCount + i], false);
+  }
+  return 0;
+}
+
+int main() {
+  pthread_t t;
+  pthread_create(&t, 0, Thread, 0);
+  sleep(1);
+  for (int i = 0; i < kTestCount; i++) {
+    fprintf(stderr, "Test %d\n", i);
+    Test(i, &atomics[i], true);
+  }
+  for (int i = 0; i < kTestCount; i++) {
+    Test(i, &atomics[kTestCount + i], true);
+  }
+  pthread_join(t, 0);
+}
+
+// CHECK-NOT: ThreadSanitizer: data race
diff --git a/lib/tsan/lit_tests/atomic_race.cc b/lib/tsan/lit_tests/atomic_race.cc
new file mode 100644
index 0000000000000..360b81238889d
--- /dev/null
+++ b/lib/tsan/lit_tests/atomic_race.cc
@@ -0,0 +1,80 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <unistd.h>
+#include <stdio.h>
+
+const int kTestCount = 4;
+typedef long long T;
+T atomics[kTestCount * 2];
+
+void Test(int test, T *p, bool main_thread) {
+  volatile T sink;
+  if (test == 0) {
+    if (main_thread)
+      __atomic_fetch_add(p, 1, __ATOMIC_RELAXED);
+    else
+      *p = 42;
+  } else if (test == 1) {
+    if (main_thread)
+      __atomic_fetch_add(p, 1, __ATOMIC_RELAXED);
+    else
+      sink = *p;
+  } else if (test == 2) {
+    if (main_thread)
+      sink = __atomic_load_n(p, __ATOMIC_SEQ_CST);
+    else
+      *p = 42;
+  } else if (test == 3) {
+    if (main_thread)
+      __atomic_store_n(p, 1, __ATOMIC_SEQ_CST);
+    else
+      sink = *p;
+  }
+}
+
+void *Thread(void *p) {
+  for (int i = 0; i < kTestCount; i++) {
+    Test(i, &atomics[i], false);
+  }
+  sleep(2);
+  for (int i = 0; i < kTestCount; i++) {
+    fprintf(stderr, "Test %d reverse\n", i);
+    Test(i, &atomics[kTestCount + i], false);
+  }
+  return 0;
+}
+
+int main() {
+  pthread_t t;
+  pthread_create(&t, 0, Thread, 0);
+  sleep(1);
+  for (int i = 0; i < kTestCount; i++) {
+    fprintf(stderr, "Test %d\n", i);
+    Test(i, &atomics[i], true);
+  }
+  for (int i = 0; i < kTestCount; i++) {
+    Test(i, &atomics[kTestCount + i], true);
+  }
+  pthread_join(t, 0);
+}
+
+// CHECK: Test 0
+// CHECK: ThreadSanitizer: data race
+// CHECK-NOT: SUMMARY{{.*}}tsan_interface_atomic
+// CHECK: Test 1
+// CHECK: ThreadSanitizer: data race
+// CHECK-NOT: SUMMARY{{.*}}tsan_interface_atomic
+// CHECK: Test 2
+// CHECK: ThreadSanitizer: data race
+// CHECK-NOT: SUMMARY{{.*}}tsan_interface_atomic
+// CHECK: Test 3
+// CHECK: ThreadSanitizer: data race
+// CHECK-NOT: SUMMARY{{.*}}tsan_interface_atomic
+// CHECK: Test 0 reverse
+// CHECK: ThreadSanitizer: data race
+// CHECK: Test 1 reverse
+// CHECK: ThreadSanitizer: data race
+// CHECK: Test 2 reverse
+// CHECK: ThreadSanitizer: data race
+// CHECK: Test 3 reverse
+// CHECK: ThreadSanitizer: data race
diff --git a/lib/tsan/lit_tests/atomic_stack.cc b/lib/tsan/lit_tests/atomic_stack.cc
new file mode 100644
index 0000000000000..50f6a8a889ca8
--- /dev/null
+++ b/lib/tsan/lit_tests/atomic_stack.cc
@@ -0,0 +1,29 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <unistd.h>
+
+int Global;
+
+void *Thread1(void *x) {
+  sleep(1);
+  __atomic_fetch_add(&Global, 1, __ATOMIC_RELAXED);
+  return NULL;
+}
+
+void *Thread2(void *x) {
+  Global++;
+  return NULL;
+}
+
+int main() {
+  pthread_t t[2];
+  pthread_create(&t[0], NULL, Thread1, NULL);
+  pthread_create(&t[1], NULL, Thread2, NULL);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+}
+
+// CHECK: WARNING: ThreadSanitizer: data race
+// CHECK:   Atomic write of size 4
+// CHECK:     #0 __tsan_atomic32_fetch_add
+// CHECK:     #1 Thread1
diff --git a/lib/tsan/lit_tests/benign_race.cc b/lib/tsan/lit_tests/benign_race.cc
new file mode 100644
index 0000000000000..a4d4d23c362af
--- /dev/null
+++ b/lib/tsan/lit_tests/benign_race.cc
@@ -0,0 +1,39 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <unistd.h>
+
+int Global;
+int WTFGlobal;
+
+extern "C" {
+void AnnotateBenignRaceSized(const char *f, int l,
+                             void *mem, unsigned int size, const char *desc);
+void WTFAnnotateBenignRaceSized(const char *f, int l,
+                                void *mem, unsigned int size,
+                                const char *desc);
+}
+
+
+void *Thread(void *x) {
+  Global = 42;
+  WTFGlobal = 142;
+  return 0;
+}
+
+int main() {
+  AnnotateBenignRaceSized(__FILE__, __LINE__,
+                          &Global, sizeof(Global), "Race on Global");
+  WTFAnnotateBenignRaceSized(__FILE__, __LINE__,
+                             &WTFGlobal, sizeof(WTFGlobal),
+                             "Race on WTFGlobal");
+  pthread_t t;
+  pthread_create(&t, 0, Thread, 0);
+  sleep(1);
+  Global = 43;
+  WTFGlobal = 143;
+  pthread_join(t, 0);
+  printf("OK\n");
+}
+
+// CHECK-NOT: WARNING: ThreadSanitizer: data race
diff --git a/lib/tsan/lit_tests/free_race.c b/lib/tsan/lit_tests/free_race.c
index 7a2ec0cdbed0c..ff71a4d2116b4 100644
--- a/lib/tsan/lit_tests/free_race.c
+++ b/lib/tsan/lit_tests/free_race.c
@@ -40,4 +40,5 @@ int main() {
 // CHECK:     #1 main
 // CHECK:   Previous write of size 8 at {{.*}} by thread T1{{.*}}:
 // CHECK:     #0 free
-// CHECK:     #1 Thread1
+// CHECK:     #{{(1|2)}} Thread1
+// CHECK: SUMMARY: ThreadSanitizer: heap-use-after-free{{.*}}Thread2
diff --git a/lib/tsan/lit_tests/free_race2.c b/lib/tsan/lit_tests/free_race2.c
index 095f82ea08188..f20774b2d8d4a 100644
--- a/lib/tsan/lit_tests/free_race2.c
+++ b/lib/tsan/lit_tests/free_race2.c
@@ -22,5 +22,5 @@ int main() {
 // CHECK:     #1 main
 // CHECK:   Previous write of size 8 at {{.*}} by main thread:
 // CHECK:     #0 free
-// CHECK:     #1 foo
-// CHECK:     #2 main
+// CHECK:     #{{1|2}} foo
+// CHECK:     #{{2|3}} main
diff --git a/lib/tsan/lit_tests/inlined_memcpy_race.cc b/lib/tsan/lit_tests/inlined_memcpy_race.cc
new file mode 100644
index 0000000000000..6efe5a956e9db
--- /dev/null
+++ b/lib/tsan/lit_tests/inlined_memcpy_race.cc
@@ -0,0 +1,55 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
+
+int x[4], y[4], z[4];
+
+void *MemCpyThread(void *a) {
+  memcpy((int*)a, z, 16);
+  return NULL;
+}
+
+void *MemMoveThread(void *a) {
+  memmove((int*)a, z, 16);
+  return NULL;
+}
+
+void *MemSetThread(void *a) {
+  sleep(1);
+  memset((int*)a, 0, 16);
+  return NULL;
+}
+
+int main() {
+  pthread_t t[2];
+  // Race on x between memcpy and memset
+  pthread_create(&t[0], NULL, MemCpyThread, x);
+  pthread_create(&t[1], NULL, MemSetThread, x);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+  // Race on y between memmove and memset
+  pthread_create(&t[0], NULL, MemMoveThread, y);
+  pthread_create(&t[1], NULL, MemSetThread, y);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+
+  printf("PASS\n");
+  return 0;
+}
+
+// CHECK: WARNING: ThreadSanitizer: data race
+// CHECK:   #0 memset
+// CHECK:   #1 MemSetThread
+// CHECK:  Previous write
+// CHECK:   #0 memcpy
+// CHECK:   #1 MemCpyThread
+
+// CHECK: WARNING: ThreadSanitizer: data race
+// CHECK:   #0 memset
+// CHECK:   #1 MemSetThread
+// CHECK:  Previous write
+// CHECK:   #0 memmove
+// CHECK:   #1 MemMoveThread
diff --git a/lib/tsan/lit_tests/java.h b/lib/tsan/lit_tests/java.h
index 7d61f58028645..04094197edb7d 100644
--- a/lib/tsan/lit_tests/java.h
+++ b/lib/tsan/lit_tests/java.h
@@ -14,4 +14,6 @@ void __tsan_java_mutex_lock(jptr addr);
 void __tsan_java_mutex_unlock(jptr addr);
 void __tsan_java_mutex_read_lock(jptr addr);
 void __tsan_java_mutex_read_unlock(jptr addr);
+void __tsan_java_mutex_lock_rec(jptr addr, int rec);
+int  __tsan_java_mutex_unlock_rec(jptr addr);
 }
diff --git a/lib/tsan/lit_tests/java_lock.cc b/lib/tsan/lit_tests/java_lock.cc
index f66f1e7097fab..d9db103504de1 100644
--- a/lib/tsan/lit_tests/java_lock.cc
+++ b/lib/tsan/lit_tests/java_lock.cc
@@ -1,10 +1,12 @@
 // RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
 #include "java.h"
+#include <unistd.h>
 
 jptr varaddr;
 jptr lockaddr;
 
 void *Thread(void *p) {
+  sleep(1);
   __tsan_java_mutex_lock(lockaddr);
   *(int*)varaddr = 42;
   __tsan_java_mutex_unlock(lockaddr);
diff --git a/lib/tsan/lit_tests/java_lock_rec.cc b/lib/tsan/lit_tests/java_lock_rec.cc
new file mode 100644
index 0000000000000..5cc80d4a33eff
--- /dev/null
+++ b/lib/tsan/lit_tests/java_lock_rec.cc
@@ -0,0 +1,54 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include "java.h"
+#include <unistd.h>
+
+jptr varaddr;
+jptr lockaddr;
+
+void *Thread(void *p) {
+  __tsan_java_mutex_lock(lockaddr);
+  __tsan_java_mutex_lock(lockaddr);
+  *(int*)varaddr = 42;
+  int rec = __tsan_java_mutex_unlock_rec(lockaddr);
+  if (rec != 2) {
+    printf("FAILED 0 rec=%d\n", rec);
+    exit(1);
+  }
+  sleep(2);
+  __tsan_java_mutex_lock_rec(lockaddr, rec);
+  if (*(int*)varaddr != 43) {
+    printf("FAILED 3 var=%d\n", *(int*)varaddr);
+    exit(1);
+  }
+  __tsan_java_mutex_unlock(lockaddr);
+  __tsan_java_mutex_unlock(lockaddr);
+  return 0;
+}
+
+int main() {
+  int const kHeapSize = 1024 * 1024;
+  void *jheap = malloc(kHeapSize);
+  __tsan_java_init((jptr)jheap, kHeapSize);
+  const int kBlockSize = 16;
+  __tsan_java_alloc((jptr)jheap, kBlockSize);
+  varaddr = (jptr)jheap;
+  *(int*)varaddr = 0;
+  lockaddr = (jptr)jheap + 8;
+  pthread_t th;
+  pthread_create(&th, 0, Thread, 0);
+  sleep(1);
+  __tsan_java_mutex_lock(lockaddr);
+  if (*(int*)varaddr != 42) {
+    printf("FAILED 1 var=%d\n", *(int*)varaddr);
+    exit(1);
+  }
+  *(int*)varaddr = 43;
+  __tsan_java_mutex_unlock(lockaddr);
+  pthread_join(th, 0);
+  __tsan_java_free((jptr)jheap, kBlockSize);
+  printf("OK\n");
+  return __tsan_java_fini();
+}
+
+// CHECK-NOT: WARNING: ThreadSanitizer: data race
+// CHECK-NOT: FAILED
diff --git a/lib/tsan/lit_tests/java_lock_rec_race.cc b/lib/tsan/lit_tests/java_lock_rec_race.cc
new file mode 100644
index 0000000000000..61626aaddc0d3
--- /dev/null
+++ b/lib/tsan/lit_tests/java_lock_rec_race.cc
@@ -0,0 +1,48 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include "java.h"
+#include <unistd.h>
+
+jptr varaddr;
+jptr lockaddr;
+
+void *Thread(void *p) {
+  __tsan_java_mutex_lock(lockaddr);
+  __tsan_java_mutex_lock(lockaddr);
+  __tsan_java_mutex_lock(lockaddr);
+  int rec = __tsan_java_mutex_unlock_rec(lockaddr);
+  if (rec != 3) {
+    printf("FAILED 0 rec=%d\n", rec);
+    exit(1);
+  }
+  *(int*)varaddr = 42;
+  sleep(2);
+  __tsan_java_mutex_lock_rec(lockaddr, rec);
+  __tsan_java_mutex_unlock(lockaddr);
+  __tsan_java_mutex_unlock(lockaddr);
+  __tsan_java_mutex_unlock(lockaddr);
+  return 0;
+}
+
+int main() {
+  int const kHeapSize = 1024 * 1024;
+  void *jheap = malloc(kHeapSize);
+  __tsan_java_init((jptr)jheap, kHeapSize);
+  const int kBlockSize = 16;
+  __tsan_java_alloc((jptr)jheap, kBlockSize);
+  varaddr = (jptr)jheap;
+  *(int*)varaddr = 0;
+  lockaddr = (jptr)jheap + 8;
+  pthread_t th;
+  pthread_create(&th, 0, Thread, 0);
+  sleep(1);
+  __tsan_java_mutex_lock(lockaddr);
+  *(int*)varaddr = 43;
+  __tsan_java_mutex_unlock(lockaddr);
+  pthread_join(th, 0);
+  __tsan_java_free((jptr)jheap, kBlockSize);
+  printf("OK\n");
+  return __tsan_java_fini();
+}
+
+// CHECK: WARNING: ThreadSanitizer: data race
+// CHECK-NOT: FAILED
diff --git a/lib/tsan/lit_tests/java_rwlock.cc b/lib/tsan/lit_tests/java_rwlock.cc
index 1e8940afd7d08..d1f38733ba035 100644
--- a/lib/tsan/lit_tests/java_rwlock.cc
+++ b/lib/tsan/lit_tests/java_rwlock.cc
@@ -1,10 +1,12 @@
 // RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
 #include "java.h"
+#include <unistd.h>
 
 jptr varaddr;
 jptr lockaddr;
 
 void *Thread(void *p) {
+  sleep(1);
   __tsan_java_mutex_read_lock(lockaddr);
   *(int*)varaddr = 42;
   __tsan_java_mutex_read_unlock(lockaddr);
diff --git a/lib/tsan/lit_tests/lit.cfg b/lib/tsan/lit_tests/lit.cfg
index 7e2db7b8fd0b0..d483d2fcbdc60 100644
--- a/lib/tsan/lit_tests/lit.cfg
+++ b/lib/tsan/lit_tests/lit.cfg
@@ -2,6 +2,14 @@
 
 import os
 
+def get_required_attr(config, attr_name):
+  attr_value = getattr(config, attr_name, None)
+  if not attr_value:
+    lit.fatal("No attribute %r in test configuration! You may need to run "
+              "tests from your build directory or add this attribute "
+              "to lit.site.cfg " % attr_name)
+  return attr_value
+
 # Setup config name.
 config.name = 'ThreadSanitizer'
 
@@ -30,14 +38,6 @@ if llvm_src_root is None:
   if not llvm_config:
     DisplayNoConfigMessage()
 
-  # Validate that llvm-config points to the same source tree.
-  llvm_src_root = lit.util.capture(["llvm-config", "--src-root"]).strip()
-  tsan_test_src_root = os.path.join(llvm_src_root, "projects", "compiler-rt",
-                                    "lib", "tsan", "lit_tests")
-  if (os.path.realpath(tsan_test_src_root) !=
-      os.path.realpath(config.test_source_root)):
-    DisplayNoConfigMessage()
-
   # Find out the presumed location of generated site config.
   llvm_obj_root = lit.util.capture(["llvm-config", "--obj-root"]).strip()
   tsan_site_cfg = os.path.join(llvm_obj_root, "projects", "compiler-rt",
@@ -49,8 +49,9 @@ if llvm_src_root is None:
   raise SystemExit
 
 # Setup attributes common for all compiler-rt projects.
-compiler_rt_lit_cfg = os.path.join(llvm_src_root, "projects", "compiler-rt",
-                                   "lib", "lit.common.cfg")
+compiler_rt_src_root = get_required_attr(config, 'compiler_rt_src_root')
+compiler_rt_lit_cfg = os.path.join(compiler_rt_src_root, "lib",
+                                   "lit.common.cfg")
 if (not compiler_rt_lit_cfg) or (not os.path.exists(compiler_rt_lit_cfg)):
   lit.fatal("Can't find common compiler-rt lit config at: %r"
             % compiler_rt_lit_cfg)
@@ -69,11 +70,8 @@ config.environment['TSAN_OPTIONS'] = tsan_options
 # Setup default compiler flags used with -fsanitize=thread option.
 # FIXME: Review the set of required flags and check if it can be reduced.
 clang_tsan_cflags = ("-fsanitize=thread "
-                      + "-fPIE "
-                      + "-fno-builtin "
                       + "-g "
                       + "-Wall "
-                      + "-pie "
                       + "-lpthread "
                       + "-ldl ")
 clang_tsan_cxxflags = "-ccc-cxx " + clang_tsan_cflags
diff --git a/lib/tsan/lit_tests/lit.site.cfg.in b/lib/tsan/lit_tests/lit.site.cfg.in
index b1c6ccf544ea5..07b521af061f7 100644
--- a/lib/tsan/lit_tests/lit.site.cfg.in
+++ b/lib/tsan/lit_tests/lit.site.cfg.in
@@ -4,6 +4,7 @@
 config.clang = "@LLVM_BINARY_DIR@/bin/clang"
 config.host_os = "@HOST_OS@"
 config.llvm_src_root = "@LLVM_SOURCE_DIR@"
+config.compiler_rt_src_root = "@COMPILER_RT_SOURCE_DIR@"
 config.llvm_tools_dir = "@LLVM_TOOLS_DIR@"
 config.target_triple = "@TARGET_TRIPLE@"
 
diff --git a/lib/tsan/lit_tests/load_shared_lib.cc b/lib/tsan/lit_tests/load_shared_lib.cc
new file mode 100644
index 0000000000000..dd6fa0964f4af
--- /dev/null
+++ b/lib/tsan/lit_tests/load_shared_lib.cc
@@ -0,0 +1,44 @@
+// Check that if the list of shared libraries changes between the two race
+// reports, the second report occurring in a new shared library is still
+// symbolized correctly.
+
+// RUN: %clangxx_tsan -O1 %p/SharedLibs/load_shared_lib-so.cc \
+// RUN:     -fPIC -shared -o %t-so.so
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+
+#include <dlfcn.h>
+#include <pthread.h>
+#include <stdio.h>
+
+#include <string>
+
+int GLOB = 0;
+
+void *write_glob(void *unused) {
+  GLOB++;
+  return NULL;
+}
+
+void race_two_threads(void *(*access_callback)(void *unused)) {
+  pthread_t t1, t2;
+  pthread_create(&t1, NULL, access_callback, NULL);
+  pthread_create(&t2, NULL, access_callback, NULL);
+  pthread_join(t1, NULL);
+  pthread_join(t2, NULL);
+}
+
+int main(int argc, char *argv[]) {
+  std::string path = std::string(argv[0]) + std::string("-so.so");
+  race_two_threads(write_glob);
+  // CHECK: write_glob
+  void *lib = dlopen(path.c_str(), RTLD_NOW);
+    if (!lib) {
+    printf("error in dlopen(): %s\n", dlerror());
+    return 1;
+  }
+  void *(*write_from_so)(void *unused);
+  *(void **)&write_from_so = dlsym(lib, "write_from_so");
+  race_two_threads(write_from_so);
+  // CHECK: write_from_so
+  return 0;
+}
diff --git a/lib/tsan/lit_tests/longjmp.cc b/lib/tsan/lit_tests/longjmp.cc
new file mode 100644
index 0000000000000..d9ca4ca5e6e9d
--- /dev/null
+++ b/lib/tsan/lit_tests/longjmp.cc
@@ -0,0 +1,22 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <stdio.h>
+#include <stdlib.h>
+#include <setjmp.h>
+
+int foo(jmp_buf env) {
+  longjmp(env, 42);
+}
+
+int main() {
+  jmp_buf env;
+  if (setjmp(env) == 42) {
+    printf("JUMPED\n");
+    return 0;
+  }
+  foo(env);
+  printf("FAILED\n");
+  return 0;
+}
+
+// CHECK-NOT: FAILED
+// CHECK: JUMPED
diff --git a/lib/tsan/lit_tests/longjmp2.cc b/lib/tsan/lit_tests/longjmp2.cc
new file mode 100644
index 0000000000000..0d551fa19d949
--- /dev/null
+++ b/lib/tsan/lit_tests/longjmp2.cc
@@ -0,0 +1,24 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <stdio.h>
+#include <stdlib.h>
+#include <setjmp.h>
+
+int foo(sigjmp_buf env) {
+  printf("env=%p\n", env);
+  siglongjmp(env, 42);
+}
+
+int main() {
+  sigjmp_buf env;
+  printf("env=%p\n", env);
+  if (sigsetjmp(env, 1) == 42) {
+    printf("JUMPED\n");
+    return 0;
+  }
+  foo(env);
+  printf("FAILED\n");
+  return 0;
+}
+
+// CHECK-NOT: FAILED
+// CHECK: JUMPED
diff --git a/lib/tsan/lit_tests/longjmp3.cc b/lib/tsan/lit_tests/longjmp3.cc
new file mode 100644
index 0000000000000..87fabd0b3be27
--- /dev/null
+++ b/lib/tsan/lit_tests/longjmp3.cc
@@ -0,0 +1,48 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <setjmp.h>
+
+void bar(jmp_buf env) {
+  volatile int x = 42;
+  longjmp(env, 42);
+  x++;
+}
+
+void foo(jmp_buf env) {
+  volatile int x = 42;
+  bar(env);
+  x++;
+}
+
+void badguy() {
+  pthread_mutex_t mtx;
+  pthread_mutex_init(&mtx, 0);
+  pthread_mutex_lock(&mtx);
+  pthread_mutex_destroy(&mtx);
+}
+
+void mymain() {
+  jmp_buf env;
+  if (setjmp(env) == 42) {
+    badguy();
+    return;
+  }
+  foo(env);
+  printf("FAILED\n");
+}
+
+int main() {
+  volatile int x = 42;
+  mymain();
+  return x;
+}
+
+// CHECK-NOT: FAILED
+// CHECK: WARNING: ThreadSanitizer: destroy of a locked mutex
+// CHECK:   #0 pthread_mutex_destroy
+// CHECK:   #1 badguy
+// CHECK:   #2 mymain
+// CHECK:   #3 main
+
diff --git a/lib/tsan/lit_tests/longjmp4.cc b/lib/tsan/lit_tests/longjmp4.cc
new file mode 100644
index 0000000000000..a8764dda5a6bd
--- /dev/null
+++ b/lib/tsan/lit_tests/longjmp4.cc
@@ -0,0 +1,51 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <setjmp.h>
+#include <string.h>
+
+void bar(jmp_buf env) {
+  volatile int x = 42;
+  jmp_buf env2;
+  memcpy(env2, env, sizeof(jmp_buf));
+  longjmp(env2, 42);
+  x++;
+}
+
+void foo(jmp_buf env) {
+  volatile int x = 42;
+  bar(env);
+  x++;
+}
+
+void badguy() {
+  pthread_mutex_t mtx;
+  pthread_mutex_init(&mtx, 0);
+  pthread_mutex_lock(&mtx);
+  pthread_mutex_destroy(&mtx);
+}
+
+void mymain() {
+  jmp_buf env;
+  if (setjmp(env) == 42) {
+    badguy();
+    return;
+  }
+  foo(env);
+  printf("FAILED\n");
+}
+
+int main() {
+  volatile int x = 42;
+  mymain();
+  return x;
+}
+
+// CHECK-NOT: FAILED
+// CHECK: WARNING: ThreadSanitizer: destroy of a locked mutex
+// CHECK:   #0 pthread_mutex_destroy
+// CHECK:   #1 badguy
+// CHECK:   #2 mymain
+// CHECK:   #3 main
+
diff --git a/lib/tsan/lit_tests/malloc_overflow.cc b/lib/tsan/lit_tests/malloc_overflow.cc
new file mode 100644
index 0000000000000..19423c5f93f14
--- /dev/null
+++ b/lib/tsan/lit_tests/malloc_overflow.cc
@@ -0,0 +1,22 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <stdio.h>
+#include <stdlib.h>
+
+int main() {
+  void *p = malloc((size_t)-1);
+  if (p != 0)
+    printf("FAIL malloc(-1) = %p\n", p);
+  p = malloc((size_t)-1 / 2);
+  if (p != 0)
+    printf("FAIL malloc(-1/2) = %p\n", p);
+  p = calloc((size_t)-1, (size_t)-1);
+  if (p != 0)
+    printf("FAIL calloc(-1, -1) = %p\n", p);
+  p = calloc((size_t)-1 / 2, (size_t)-1 / 2);
+  if (p != 0)
+    printf("FAIL calloc(-1/2, -1/2) = %p\n", p);
+  printf("OK\n");
+}
+
+// CHECK-NOT: FAIL
+// CHECK-NOT: failed to allocate
diff --git a/lib/tsan/lit_tests/malloc_stack.cc b/lib/tsan/lit_tests/malloc_stack.cc
new file mode 100644
index 0000000000000..c185623ff5ca0
--- /dev/null
+++ b/lib/tsan/lit_tests/malloc_stack.cc
@@ -0,0 +1,25 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <unistd.h>
+
+_Atomic(int*) p;
+
+void *thr(void *a) {
+  sleep(1);
+  int *pp = __c11_atomic_load(&p, __ATOMIC_RELAXED);
+  *pp = 42;
+  return 0;
+}
+
+int main() {
+  pthread_t th;
+  pthread_create(&th, 0, thr, p);
+  __c11_atomic_store(&p, new int, __ATOMIC_RELAXED);
+  pthread_join(th, 0);
+}
+
+// CHECK: data race
+// CHECK:   Previous write
+// CHECK:     #0 operator new
+// CHECK:   Location is heap block
+// CHECK:     #0 operator new
diff --git a/lib/tsan/lit_tests/memcpy_race.cc b/lib/tsan/lit_tests/memcpy_race.cc
index 806740dda2418..857728ba05400 100644
--- a/lib/tsan/lit_tests/memcpy_race.cc
+++ b/lib/tsan/lit_tests/memcpy_race.cc
@@ -10,13 +10,15 @@ char *data1 = new char[10];
 char *data2 = new char[10];
 
 void *Thread1(void *x) {
-  memcpy(data+5, data1, 1);
+  static volatile int size = 1;
+  memcpy(data+5, data1, size);
   return NULL;
 }
 
 void *Thread2(void *x) {
+  static volatile int size = 4;
   sleep(1);
-  memcpy(data+3, data2, 4);
+  memcpy(data+3, data2, size);
   return NULL;
 }
 
diff --git a/lib/tsan/lit_tests/mutex_destroy_locked.cc b/lib/tsan/lit_tests/mutex_destroy_locked.cc
index 991eaf5426e22..27a04248b172b 100644
--- a/lib/tsan/lit_tests/mutex_destroy_locked.cc
+++ b/lib/tsan/lit_tests/mutex_destroy_locked.cc
@@ -19,3 +19,4 @@ int main() {
 // CHECK:   Mutex {{.*}} created at:
 // CHECK:     #0 pthread_mutex_init
 // CHECK:     #1 main
+// CHECK: SUMMARY: ThreadSanitizer: destroy of a locked mutex{{.*}}main
diff --git a/lib/tsan/lit_tests/mutexset7.cc b/lib/tsan/lit_tests/mutexset7.cc
index 141bde2b5015e..3ec1b52029836 100644
--- a/lib/tsan/lit_tests/mutexset7.cc
+++ b/lib/tsan/lit_tests/mutexset7.cc
@@ -4,6 +4,7 @@
 #include <unistd.h>
 
 int Global;
+__thread int huge[1024*1024];
 
 void *Thread1(void *x) {
   sleep(1);
diff --git a/lib/tsan/lit_tests/mutexset8.cc b/lib/tsan/lit_tests/mutexset8.cc
new file mode 100644
index 0000000000000..6db63f7d16dbd
--- /dev/null
+++ b/lib/tsan/lit_tests/mutexset8.cc
@@ -0,0 +1,39 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <unistd.h>
+
+int Global;
+pthread_mutex_t *mtx;
+
+void *Thread1(void *x) {
+  sleep(1);
+  pthread_mutex_lock(mtx);
+  Global++;
+  pthread_mutex_unlock(mtx);
+  return NULL;
+}
+
+void *Thread2(void *x) {
+  Global--;
+  return NULL;
+}
+
+int main() {
+  // CHECK: WARNING: ThreadSanitizer: data race
+  // CHECK:   Write of size 4 at {{.*}} by thread T1
+  // CHECK:                         (mutexes: write [[M1:M[0-9]+]]):
+  // CHECK:   Previous write of size 4 at {{.*}} by thread T2:
+  // CHECK:   Mutex [[M1]] created at:
+  // CHECK:     #0 pthread_mutex_init
+  // CHECK:     #1 main {{.*}}/mutexset8.cc
+  mtx = new pthread_mutex_t;
+  pthread_mutex_init(mtx, 0);
+  pthread_t t[2];
+  pthread_create(&t[0], NULL, Thread1, NULL);
+  pthread_create(&t[1], NULL, Thread2, NULL);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+  pthread_mutex_destroy(mtx);
+  delete mtx;
+}
diff --git a/lib/tsan/lit_tests/oob_race.cc b/lib/tsan/lit_tests/oob_race.cc
new file mode 100644
index 0000000000000..2e7f0593fd8d9
--- /dev/null
+++ b/lib/tsan/lit_tests/oob_race.cc
@@ -0,0 +1,24 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+
+const long kOffset = 64*1024;
+
+void *Thread(void *p) {
+  ((char*)p)[-kOffset] = 43;
+  return 0;
+}
+
+int main() {
+  char *volatile p0 = new char[16];
+  delete[] p0;
+  char *p = new char[32];
+  pthread_t th;
+  pthread_create(&th, 0, Thread, p);
+  p[-kOffset] = 42;
+  pthread_join(th, 0);
+}
+
+// Used to crash with CHECK failed.
+// CHECK: WARNING: ThreadSanitizer: data race
+
diff --git a/lib/tsan/lit_tests/race_on_heap.cc b/lib/tsan/lit_tests/race_on_heap.cc
index dc679e8bf3f96..35434eac18506 100644
--- a/lib/tsan/lit_tests/race_on_heap.cc
+++ b/lib/tsan/lit_tests/race_on_heap.cc
@@ -39,8 +39,8 @@ int main() {
 // ...
 // CHECK: Location is heap block of size 99 at [[ADDR]] allocated by thread T1:
 // CHCEKL     #0 malloc
-// CHECK:     #1 alloc
-// CHECK:     #2 AllocThread
+// CHECK:     #{{1|2}} alloc
+// CHECK:     #{{2|3}} AllocThread
 // ...
 // CHECK:   Thread T1 (tid={{.*}}, finished) created by main thread at:
 // CHECK:     #0 pthread_create
diff --git a/lib/tsan/lit_tests/race_on_mutex.c b/lib/tsan/lit_tests/race_on_mutex.c
index de1c2d4160a60..aff32f9bb1a2d 100644
--- a/lib/tsan/lit_tests/race_on_mutex.c
+++ b/lib/tsan/lit_tests/race_on_mutex.c
@@ -34,7 +34,7 @@ int main() {
 }
 
 // CHECK:      WARNING: ThreadSanitizer: data race
-// CHECK-NEXT:   Read of size 1 at {{.*}} by thread T2:
+// CHECK-NEXT:   Atomic read of size 1 at {{.*}} by thread T2:
 // CHECK-NEXT:     #0 pthread_mutex_lock
 // CHECK-NEXT:     #1 Thread2{{.*}} {{.*}}race_on_mutex.c:20{{(:3)?}} ({{.*}})
 // CHECK:        Previous write of size 1 at {{.*}} by thread T1:
diff --git a/lib/tsan/lit_tests/race_on_mutex2.c b/lib/tsan/lit_tests/race_on_mutex2.c
new file mode 100644
index 0000000000000..84bef75a34499
--- /dev/null
+++ b/lib/tsan/lit_tests/race_on_mutex2.c
@@ -0,0 +1,24 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <stddef.h>
+#include <unistd.h>
+
+void *Thread(void *x) {
+  pthread_mutex_lock((pthread_mutex_t*)x);
+  pthread_mutex_unlock((pthread_mutex_t*)x);
+  return 0;
+}
+
+int main() {
+  pthread_mutex_t Mtx;
+  pthread_mutex_init(&Mtx, 0);
+  pthread_t t;
+  pthread_create(&t, 0, Thread, &Mtx);
+  sleep(1);
+  pthread_mutex_destroy(&Mtx);
+  pthread_join(t, 0);
+  return 0;
+}
+
+// CHECK:      WARNING: ThreadSanitizer: data race
diff --git a/lib/tsan/lit_tests/race_on_write.cc b/lib/tsan/lit_tests/race_on_write.cc
new file mode 100644
index 0000000000000..f1b0bb1cbd6ee
--- /dev/null
+++ b/lib/tsan/lit_tests/race_on_write.cc
@@ -0,0 +1,39 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+
+int fd;
+char buf;
+
+void *Thread1(void *x) {
+  buf = 1;
+  sleep(1);
+  return NULL;
+}
+
+void *Thread2(void *x) {
+  write(fd, &buf, 1);
+  return NULL;
+}
+
+int main() {
+  fd = open("/dev/null", O_WRONLY);
+  if (fd < 0) return 1;
+  pthread_t t[2];
+  pthread_create(&t[0], NULL, Thread1, NULL);
+  sleep(1);
+  pthread_create(&t[1], NULL, Thread2, NULL);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+  close(fd);
+}
+
+// CHECK: WARNING: ThreadSanitizer: data race
+// CHECK:   Read of size 1
+// CHECK:     #0 write
+// CHECK:   Previous write of size 1
+// CHECK:     #0 Thread1
diff --git a/lib/tsan/lit_tests/signal_errno.cc b/lib/tsan/lit_tests/signal_errno.cc
index af9ccce9045a7..8181555f6f639 100644
--- a/lib/tsan/lit_tests/signal_errno.cc
+++ b/lib/tsan/lit_tests/signal_errno.cc
@@ -10,7 +10,7 @@
 pthread_t mainth;
 volatile int done;
 
-static void handler(int, siginfo_t *s, void *c) {
+static void MyHandler(int, siginfo_t *s, void *c) {
   errno = 1;
   done = 1;
 }
@@ -23,7 +23,7 @@ static void* sendsignal(void *p) {
 int main() {
   mainth = pthread_self();
   struct sigaction act = {};
-  act.sa_sigaction = &handler;
+  act.sa_sigaction = &MyHandler;
   sigaction(SIGPROF, &act, 0);
   pthread_t th;
   pthread_create(&th, 0, sendsignal, 0);
@@ -38,5 +38,6 @@ int main() {
 }
 
 // CHECK: WARNING: ThreadSanitizer: signal handler spoils errno
-// CHECK:     #0 handler(int, siginfo*, void*) {{.*}}signal_errno.cc
+// CHECK:     #0 MyHandler(int, siginfo{{(_t)?}}*, void*) {{.*}}signal_errno.cc
+// CHECK: SUMMARY: ThreadSanitizer: signal handler spoils errno{{.*}}MyHandler
 
diff --git a/lib/tsan/lit_tests/signal_malloc.cc b/lib/tsan/lit_tests/signal_malloc.cc
index cee997cdb7635..4dbc2f78ab17a 100644
--- a/lib/tsan/lit_tests/signal_malloc.cc
+++ b/lib/tsan/lit_tests/signal_malloc.cc
@@ -8,7 +8,8 @@
 static void handler(int, siginfo_t*, void*) {
   // CHECK: WARNING: ThreadSanitizer: signal-unsafe call inside of a signal
   // CHECK:     #0 malloc
-  // CHECK:     #1 handler(int, siginfo*, void*) {{.*}}signal_malloc.cc:[[@LINE+1]]
+  // CHECK:     #{{(1|2)}} handler(int, siginfo{{(_t)?}}*, void*) {{.*}}signal_malloc.cc:[[@LINE+2]]
+  // CHECK: SUMMARY: ThreadSanitizer: signal-unsafe call inside of a signal{{.*}}handler
   volatile char *p = (char*)malloc(1);
   p[0] = 0;
   free((void*)p);
diff --git a/lib/tsan/lit_tests/simple_race.cc b/lib/tsan/lit_tests/simple_race.cc
index ec29c92ee1a83..99cf228ac2f2b 100644
--- a/lib/tsan/lit_tests/simple_race.cc
+++ b/lib/tsan/lit_tests/simple_race.cc
@@ -23,3 +23,4 @@ int main() {
 }
 
 // CHECK: WARNING: ThreadSanitizer: data race
+// CHECK: SUMMARY: ThreadSanitizer: data race{{.*}}Thread
diff --git a/lib/tsan/lit_tests/test_output.sh b/lib/tsan/lit_tests/test_output.sh
index d21c9a797ad3f..1eedf6eb20a32 100755
--- a/lib/tsan/lit_tests/test_output.sh
+++ b/lib/tsan/lit_tests/test_output.sh
@@ -6,12 +6,13 @@ set -e # fail on any error
 ROOTDIR=$(dirname $0)/..
 BLACKLIST=$ROOTDIR/lit_tests/Helpers/blacklist.txt
 
-# Assuming clang is in path.
-CC=clang
-CXX=clang++
+# Assume clang and clang++ are in path.
+: ${CC:=clang}
+: ${CXX:=clang++}
+: ${FILECHECK:=FileCheck}
 
 # TODO: add testing for all of -O0...-O3
-CFLAGS="-fsanitize=thread -fsanitize-blacklist=$BLACKLIST -fPIE -O1 -g -fno-builtin -Wall"
+CFLAGS="-fsanitize=thread -fsanitize-blacklist=$BLACKLIST -fPIE -O1 -g -Wall"
 LDFLAGS="-pie -lpthread -ldl $ROOTDIR/rtl/libtsan.a"
 
 test_file() {
@@ -23,7 +24,7 @@ test_file() {
   $COMPILER $SRC $CFLAGS -c -o $OBJ
   $COMPILER $OBJ $LDFLAGS -o $EXE
   RES=$($EXE 2>&1 || true)
-  printf "%s\n" "$RES" | FileCheck $SRC
+  printf "%s\n" "$RES" | $FILECHECK $SRC
   if [ "$3" == "" ]; then
     rm -f $EXE $OBJ
   fi
@@ -35,6 +36,10 @@ if [ "$1" == "" ]; then
       echo SKIPPING FAILING TEST $c
       continue
     fi
+    if [[ $c == */load_shared_lib.cc ]]; then
+      echo TEST $c is not supported
+      continue
+    fi
     COMPILER=$CXX
     case $c in
       *.c) COMPILER=$CC
diff --git a/lib/tsan/lit_tests/thread_end_with_ignore.cc b/lib/tsan/lit_tests/thread_end_with_ignore.cc
new file mode 100644
index 0000000000000..960a477c5ad33
--- /dev/null
+++ b/lib/tsan/lit_tests/thread_end_with_ignore.cc
@@ -0,0 +1,19 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+
+extern "C" void AnnotateIgnoreReadsBegin(const char *f, int l);
+
+void *Thread(void *x) {
+  AnnotateIgnoreReadsBegin("", 0);
+  return 0;
+}
+
+int main() {
+  pthread_t t;
+  pthread_create(&t, 0, Thread, 0);
+  pthread_join(t, 0);
+}
+
+// CHECK: ThreadSanitizer: thread T1 finished with ignores enabled
+
diff --git a/lib/tsan/lit_tests/thread_end_with_ignore2.cc b/lib/tsan/lit_tests/thread_end_with_ignore2.cc
new file mode 100644
index 0000000000000..8f743ae2f4a4a
--- /dev/null
+++ b/lib/tsan/lit_tests/thread_end_with_ignore2.cc
@@ -0,0 +1,9 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+extern "C" void AnnotateIgnoreWritesBegin(const char *f, int l);
+
+int main() {
+  AnnotateIgnoreWritesBegin("", 0);
+}
+
+// CHECK: ThreadSanitizer: thread T0 finished with ignores enabled
+
diff --git a/lib/tsan/lit_tests/thread_leak3.c b/lib/tsan/lit_tests/thread_leak3.c
index c48219fe73fae..3577164cad4a7 100644
--- a/lib/tsan/lit_tests/thread_leak3.c
+++ b/lib/tsan/lit_tests/thread_leak3.c
@@ -1,5 +1,6 @@
 // RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
 #include <pthread.h>
+#include <unistd.h>
 
 void *Thread(void *x) {
   return 0;
@@ -8,7 +9,9 @@ void *Thread(void *x) {
 int main() {
   pthread_t t;
   pthread_create(&t, 0, Thread, 0);
+  sleep(1);
   return 0;
 }
 
 // CHECK: WARNING: ThreadSanitizer: thread leak
+// CHECK: SUMMARY: ThreadSanitizer: thread leak{{.*}}main
diff --git a/lib/tsan/lit_tests/thread_leak4.c b/lib/tsan/lit_tests/thread_leak4.c
new file mode 100644
index 0000000000000..f9fad0360d34b
--- /dev/null
+++ b/lib/tsan/lit_tests/thread_leak4.c
@@ -0,0 +1,18 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <unistd.h>
+#include <stdio.h>
+
+void *Thread(void *x) {
+  sleep(10);
+  return 0;
+}
+
+int main() {
+  pthread_t t;
+  pthread_create(&t, 0, Thread, 0);
+  printf("OK\n");
+  return 0;
+}
+
+// CHECK-NOT: WARNING: ThreadSanitizer: thread leak
diff --git a/lib/tsan/lit_tests/thread_leak5.c b/lib/tsan/lit_tests/thread_leak5.c
new file mode 100644
index 0000000000000..fc72b149ec250
--- /dev/null
+++ b/lib/tsan/lit_tests/thread_leak5.c
@@ -0,0 +1,19 @@
+// RUN: %clang_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <unistd.h>
+
+void *Thread(void *x) {
+  return 0;
+}
+
+int main() {
+  for (int i = 0; i < 5; i++) {
+    pthread_t t;
+    pthread_create(&t, 0, Thread, 0);
+  }
+  sleep(1);
+  return 0;
+}
+
+// CHECK: WARNING: ThreadSanitizer: thread leak
+// CHECK:   And 4 more similar thread leaks
diff --git a/lib/tsan/lit_tests/thread_name.cc b/lib/tsan/lit_tests/thread_name.cc
index 0ca0b17699765..37f308ffbc0cf 100644
--- a/lib/tsan/lit_tests/thread_name.cc
+++ b/lib/tsan/lit_tests/thread_name.cc
@@ -15,7 +15,11 @@ void *Thread1(void *x) {
 }
 
 void *Thread2(void *x) {
+#if SANITIZER_LINUX && __GLIBC_PREREQ(2, 12)
   pthread_setname_np(pthread_self(), "Thread2");
+#else
+  AnnotateThreadName(__FILE__, __LINE__, "Thread2");
+#endif
   Global--;
   return NULL;
 }
diff --git a/lib/tsan/lit_tests/tsan-vs-gvn.cc b/lib/tsan/lit_tests/tsan-vs-gvn.cc
new file mode 100644
index 0000000000000..40ae724b78e12
--- /dev/null
+++ b/lib/tsan/lit_tests/tsan-vs-gvn.cc
@@ -0,0 +1,38 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+// RUN: %clangxx_tsan -O2 %s -o %t && %t 2>&1 | FileCheck %s
+// RUN: %clangxx_tsan -O3 %s -o %t && %t 2>&1 | FileCheck %s
+//
+// Check that load widening is not tsan-hostile.
+#include <pthread.h>
+#include <stdio.h>
+#include <string.h>
+
+struct {
+  int i;
+  char c1, c2, c3, c4;
+} S;
+
+int G;
+
+void *Thread1(void *x) {
+  G = S.c1 + S.c3;
+  return NULL;
+}
+
+void *Thread2(void *x) {
+  S.c2 = 1;
+  return NULL;
+}
+
+int main() {
+  pthread_t t[2];
+  memset(&S, 123, sizeof(S));
+  pthread_create(&t[0], NULL, Thread1, NULL);
+  pthread_create(&t[1], NULL, Thread2, NULL);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+  printf("PASS\n");
+}
+
+// CHECK-NOT: WARNING: ThreadSanitizer: data race
+// CHECK: PASS
diff --git a/lib/tsan/lit_tests/unaligned_norace.cc b/lib/tsan/lit_tests/unaligned_norace.cc
new file mode 100644
index 0000000000000..792224b80126b
--- /dev/null
+++ b/lib/tsan/lit_tests/unaligned_norace.cc
@@ -0,0 +1,84 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+
+uint64_t objs[8*3*3*2][3];
+
+extern "C" {
+uint16_t __tsan_unaligned_read2(void *addr);
+uint32_t __tsan_unaligned_read4(void *addr);
+uint64_t __tsan_unaligned_read8(void *addr);
+void __tsan_unaligned_write2(void *addr, uint16_t v);
+void __tsan_unaligned_write4(void *addr, uint32_t v);
+void __tsan_unaligned_write8(void *addr, uint64_t v);
+}
+
+static void access(char *p, int sz, int rw) {
+  if (rw) {
+    switch (sz) {
+    case 0: __tsan_unaligned_write2(p, 0); break;
+    case 1: __tsan_unaligned_write4(p, 0); break;
+    case 2: __tsan_unaligned_write8(p, 0); break;
+    default: exit(1);
+    }
+  } else {
+    switch (sz) {
+    case 0: __tsan_unaligned_read2(p); break;
+    case 1: __tsan_unaligned_read4(p); break;
+    case 2: __tsan_unaligned_read8(p); break;
+    default: exit(1);
+    }
+  }
+}
+
+static int accesssize(int sz) {
+  switch (sz) {
+  case 0: return 2;
+  case 1: return 4;
+  case 2: return 8;
+  }
+  exit(1);
+}
+
+void Test(bool main) {
+  uint64_t *obj = objs[0];
+  for (int off = 0; off < 8; off++) {
+    for (int sz1 = 0; sz1 < 3; sz1++) {
+      for (int sz2 = 0; sz2 < 3; sz2++) {
+        for (int rw = 0; rw < 2; rw++) {
+          char *p = (char*)obj + off;
+          if (main) {
+            // printf("thr=%d off=%d sz1=%d sz2=%d rw=%d p=%p\n",
+            //        main, off, sz1, sz2, rw, p);
+            access(p, sz1, true);
+          } else {
+            p += accesssize(sz1);
+            // printf("thr=%d off=%d sz1=%d sz2=%d rw=%d p=%p\n",
+            //        main, off, sz1, sz2, rw, p);
+            access(p, sz2, rw);
+          }
+          obj += 3;
+        }
+      }
+    }
+  }
+}
+
+void *Thread(void *p) {
+  (void)p;
+  Test(false);
+  return 0;
+}
+
+int main() {
+  pthread_t th;
+  pthread_create(&th, 0, Thread, 0);
+  Test(true);
+  pthread_join(th, 0);
+  printf("OK\n");
+}
+
+// CHECK-NOT: WARNING: ThreadSanitizer:
+// CHECK: OK
diff --git a/lib/tsan/lit_tests/unaligned_race.cc b/lib/tsan/lit_tests/unaligned_race.cc
new file mode 100644
index 0000000000000..18bed8555cc5e
--- /dev/null
+++ b/lib/tsan/lit_tests/unaligned_race.cc
@@ -0,0 +1,135 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <unistd.h>
+
+uint64_t objs[8*2*(2 + 4 + 8)][2];
+
+extern "C" {
+uint16_t __sanitizer_unaligned_load16(void *addr);
+uint32_t __sanitizer_unaligned_load32(void *addr);
+uint64_t __sanitizer_unaligned_load64(void *addr);
+void __sanitizer_unaligned_store16(void *addr, uint16_t v);
+void __sanitizer_unaligned_store32(void *addr, uint32_t v);
+void __sanitizer_unaligned_store64(void *addr, uint64_t v);
+}
+
+// All this mess is to generate unique stack for each race,
+// otherwise tsan will suppress similar stacks.
+
+static void access(char *p, int sz, int rw) {
+  if (rw) {
+    switch (sz) {
+    case 0: __sanitizer_unaligned_store16(p, 0); break;
+    case 1: __sanitizer_unaligned_store32(p, 0); break;
+    case 2: __sanitizer_unaligned_store64(p, 0); break;
+    default: exit(1);
+    }
+  } else {
+    switch (sz) {
+    case 0: __sanitizer_unaligned_load16(p); break;
+    case 1: __sanitizer_unaligned_load32(p); break;
+    case 2: __sanitizer_unaligned_load64(p); break;
+    default: exit(1);
+    }
+  }
+}
+
+static int accesssize(int sz) {
+  switch (sz) {
+  case 0: return 2;
+  case 1: return 4;
+  case 2: return 8;
+  }
+  exit(1);
+}
+
+template<int off, int off2>
+static void access3(bool main, int sz1, bool rw, char *p) {
+  p += off;
+  if (main) {
+    access(p, sz1, true);
+  } else {
+    p += off2;
+    if (rw) {
+      *p = 42;
+    } else {
+       if (*p == 42)
+         printf("bingo!\n");
+    }
+  }
+}
+
+template<int off>
+static void access2(bool main, int sz1, int off2, bool rw, char *obj) {
+  if (off2 == 0)
+    access3<off, 0>(main, sz1, rw, obj);
+  else if (off2 == 1)
+    access3<off, 1>(main, sz1, rw, obj);
+  else if (off2 == 2)
+    access3<off, 2>(main, sz1, rw, obj);
+  else if (off2 == 3)
+    access3<off, 3>(main, sz1, rw, obj);
+  else if (off2 == 4)
+    access3<off, 4>(main, sz1, rw, obj);
+  else if (off2 == 5)
+    access3<off, 5>(main, sz1, rw, obj);
+  else if (off2 == 6)
+    access3<off, 6>(main, sz1, rw, obj);
+  else if (off2 == 7)
+    access3<off, 7>(main, sz1, rw, obj);
+}
+
+static void access1(bool main, int off, int sz1, int off2, bool rw, char *obj) {
+  if (off == 0)
+    access2<0>(main, sz1, off2, rw, obj);
+  else if (off == 1)
+    access2<1>(main, sz1, off2, rw, obj);
+  else if (off == 2)
+    access2<2>(main, sz1, off2, rw, obj);
+  else if (off == 3)
+    access2<3>(main, sz1, off2, rw, obj);
+  else if (off == 4)
+    access2<4>(main, sz1, off2, rw, obj);
+  else if (off == 5)
+    access2<5>(main, sz1, off2, rw, obj);
+  else if (off == 6)
+    access2<6>(main, sz1, off2, rw, obj);
+  else if (off == 7)
+    access2<7>(main, sz1, off2, rw, obj);
+}
+
+void Test(bool main) {
+  uint64_t *obj = objs[0];
+  for (int off = 0; off < 8; off++) {
+    for (int sz1 = 0; sz1 < 3; sz1++) {
+      for (int off2 = 0; off2 < accesssize(sz1); off2++) {
+        for (int rw = 0; rw < 2; rw++) {
+          // printf("thr=%d off=%d sz1=%d off2=%d rw=%d p=%p\n",
+          //        main, off, sz1, off2, rw, obj);
+          access1(main, off, sz1, off2, rw, (char*)obj);
+          obj += 2;
+        }
+      }
+    }
+  }
+}
+
+void *Thread(void *p) {
+  (void)p;
+  sleep(1);
+  Test(false);
+  return 0;
+}
+
+int main() {
+  pthread_t th;
+  pthread_create(&th, 0, Thread, 0);
+  Test(true);
+  pthread_join(th, 0);
+}
+
+// CHECK: WARNING: ThreadSanitizer: data race
+// CHECK: ThreadSanitizer: reported 224 warnings
diff --git a/lib/tsan/lit_tests/vptr_harmful_race.cc b/lib/tsan/lit_tests/vptr_harmful_race.cc
index f51ba7ee57f0b..76d31c00ad4f8 100644
--- a/lib/tsan/lit_tests/vptr_harmful_race.cc
+++ b/lib/tsan/lit_tests/vptr_harmful_race.cc
@@ -2,6 +2,7 @@
 #include <pthread.h>
 #include <semaphore.h>
 #include <stdio.h>
+#include <unistd.h>
 
 struct A {
   A() {
@@ -34,6 +35,7 @@ void *Thread1(void *x) {
 }
 
 void *Thread2(void *x) {
+  sleep(1);
   delete obj;
   return NULL;
 }
@@ -46,4 +48,4 @@ int main() {
   pthread_join(t[1], NULL);
 }
 
-// CHECK: WARNING: ThreadSanitizer: data race
+// CHECK: WARNING: ThreadSanitizer: data race on vptr
diff --git a/lib/tsan/lit_tests/vptr_harmful_race2.cc b/lib/tsan/lit_tests/vptr_harmful_race2.cc
new file mode 100644
index 0000000000000..d7e1d19a11bd3
--- /dev/null
+++ b/lib/tsan/lit_tests/vptr_harmful_race2.cc
@@ -0,0 +1,51 @@
+// RUN: %clangxx_tsan -O1 %s -o %t && %t 2>&1 | FileCheck %s
+#include <pthread.h>
+#include <semaphore.h>
+#include <stdio.h>
+#include <unistd.h>
+
+struct A {
+  A() {
+    sem_init(&sem_, 0, 0);
+  }
+  virtual void F() {
+  }
+  void Done() {
+    sem_post(&sem_);
+  }
+  virtual ~A() {
+    sem_wait(&sem_);
+    sem_destroy(&sem_);
+  }
+  sem_t sem_;
+};
+
+struct B : A {
+  virtual void F() {
+  }
+  virtual ~B() { }
+};
+
+static A *obj = new B;
+
+void *Thread1(void *x) {
+  sleep(1);
+  obj->F();
+  obj->Done();
+  return NULL;
+}
+
+void *Thread2(void *x) {
+  delete obj;
+  return NULL;
+}
+
+int main() {
+  pthread_t t[2];
+  pthread_create(&t[0], NULL, Thread1, NULL);
+  pthread_create(&t[1], NULL, Thread2, NULL);
+  pthread_join(t[0], NULL);
+  pthread_join(t[1], NULL);
+}
+
+// CHECK: WARNING: ThreadSanitizer: data race on vptr
diff --git a/lib/tsan/rtl/CMakeLists.txt b/lib/tsan/rtl/CMakeLists.txt
index d91e2e43ca4c9..f1a8ff4d65586 100644
--- a/lib/tsan/rtl/CMakeLists.txt
+++ b/lib/tsan/rtl/CMakeLists.txt
@@ -37,22 +37,15 @@ if(CAN_TARGET_x86_64 AND UNIX AND NOT APPLE)
   set(TSAN_ASM_SOURCES tsan_rtl_amd64.S)
   # Pass ASM file directly to the C++ compiler.
   set_source_files_properties(${TSAN_ASM_SOURCES} PROPERTIES
-    LANGUAGE C
-    )
-  add_library(clang_rt.tsan-x86_64 STATIC
-    ${TSAN_SOURCES}
-    ${TSAN_ASM_SOURCES}
-    $<TARGET_OBJECTS:RTInterception.x86_64>
-    $<TARGET_OBJECTS:RTSanitizerCommon.x86_64>
-    )
-  set_target_compile_flags(clang_rt.tsan-x86_64
-    ${TSAN_CFLAGS} ${TARGET_x86_64_CFLAGS}
-    )
-  list(APPEND TSAN_RUNTIME_LIBRARIES clang_rt.tsan-x86_64)
-endif()
-
-if(TSAN_RUNTIME_LIBRARIES)
-  set_property(TARGET ${TSAN_RUNTIME_LIBRARIES} APPEND PROPERTY
-    COMPILE_DEFINITIONS ${TSAN_COMMON_DEFINITIONS})
-  add_clang_compiler_rt_libraries(${TSAN_RUNTIME_LIBRARIES})
+    LANGUAGE C)
+  set(arch "x86_64")
+  add_compiler_rt_static_runtime(clang_rt.tsan-${arch} ${arch}
+    SOURCES ${TSAN_SOURCES} ${TSAN_ASM_SOURCES}
+            $<TARGET_OBJECTS:RTInterception.${arch}>
+            $<TARGET_OBJECTS:RTSanitizerCommon.${arch}>
+            $<TARGET_OBJECTS:RTSanitizerCommonLibc.${arch}>
+    CFLAGS ${TSAN_CFLAGS}
+    DEFS ${TSAN_COMMON_DEFINITIONS}
+    SYMS tsan.syms)
+  list(APPEND TSAN_RUNTIME_LIBRARIES clang_rt.tsan-${arch})
 endif()
diff --git a/lib/tsan/rtl/Makefile.mk b/lib/tsan/rtl/Makefile.mk
index a6a7fc8b86e86..2687123f731d1 100644
--- a/lib/tsan/rtl/Makefile.mk
+++ b/lib/tsan/rtl/Makefile.mk
@@ -19,7 +19,7 @@ Implementation := Generic
 # FIXME: use automatic dependencies?
 Dependencies := $(wildcard $(Dir)/*.h)
 Dependencies += $(wildcard $(Dir)/../../interception/*.h)
-Dependencies += $(wildcard $(Dir)/../../interception/mach_override/*.h)
+Dependencies += $(wildcard $(Dir)/../../sanitizer_common/*.h)
 
 # Define a convenience variable for all the tsan functions.
 TsanFunctions += $(Sources:%.cc=%) $(AsmSources:%.S=%)
diff --git a/lib/tsan/rtl/Makefile.old b/lib/tsan/rtl/Makefile.old
index f522ec6b47d7a..33944ffe96754 100644
--- a/lib/tsan/rtl/Makefile.old
+++ b/lib/tsan/rtl/Makefile.old
@@ -1,16 +1,14 @@
 CXXFLAGS = -fPIE -g -Wall -Werror -fno-builtin -DTSAN_DEBUG=$(DEBUG) -DSANITIZER_DEBUG=$(DEBUG)
+CLANG=clang
 ifeq ($(DEBUG), 0)
   CXXFLAGS += -O3
 endif
-ifeq ($(CXX), clang++)
-  CXXFLAGS+= -Wgnu
-endif
 
 # For interception. FIXME: move interception one level higher.
 INTERCEPTION=../../interception
 COMMON=../../sanitizer_common
 INCLUDES= -I../.. -I../../../include
-EXTRA_CXXFLAGS=-fno-exceptions
+EXTRA_CXXFLAGS=-fno-exceptions -fno-rtti
 NO_SYSROOT=--sysroot=.
 CXXFLAGS+=$(EXTRA_CXXFLAGS)
 CXXFLAGS+=$(CFLAGS)
@@ -21,7 +19,7 @@ ifeq ($(CXX), g++)
 endif  # CXX=g++
 endif  # DEBUG=0
 
-ifeq ($(CXX), clang++)
+ifeq ($(CXX), $(CLANG)++)
   # Global constructors are banned.
   CXXFLAGS+=-Wglobal-constructors
 endif
diff --git a/lib/tsan/rtl/tsan.syms b/lib/tsan/rtl/tsan.syms
new file mode 100644
index 0000000000000..4464a0a231c93
--- /dev/null
+++ b/lib/tsan/rtl/tsan.syms
@@ -0,0 +1,5 @@
+{
+  __tsan_*;
+  __sanitizer_syscall_pre_*;
+  __sanitizer_syscall_post_*;
+};
diff --git a/lib/tsan/rtl/tsan_defs.h b/lib/tsan/rtl/tsan_defs.h
index e0c04733f0a39..7150e2e255d80 100644
--- a/lib/tsan/rtl/tsan_defs.h
+++ b/lib/tsan/rtl/tsan_defs.h
@@ -28,16 +28,19 @@ namespace __tsan {
 const bool kGoMode = true;
 const bool kCppMode = false;
 const char *const kTsanOptionsEnv = "GORACE";
+// Go linker does not support weak symbols.
+#define CPP_WEAK
 #else
 const bool kGoMode = false;
 const bool kCppMode = true;
 const char *const kTsanOptionsEnv = "TSAN_OPTIONS";
+#define CPP_WEAK WEAK
 #endif
 
 const int kTidBits = 13;
 const unsigned kMaxTid = 1 << kTidBits;
 const unsigned kMaxTidInClock = kMaxTid * 2;  // This includes msb 'freed' bit.
-const int kClkBits = 43;
+const int kClkBits = 42;
 #ifndef TSAN_GO
 const int kShadowStackSize = 4 * 1024;
 const int kTraceStackSize = 256;
@@ -153,13 +156,13 @@ struct MD5Hash {
 MD5Hash md5_hash(const void *data, uptr size);
 
 struct ThreadState;
-struct ThreadContext;
 struct Context;
 struct ReportStack;
 class ReportDesc;
 class RegionAlloc;
 class StackTrace;
 struct MBlock;
+struct Suppression;
 
 }  // namespace __tsan
 
diff --git a/lib/tsan/rtl/tsan_fd.cc b/lib/tsan/rtl/tsan_fd.cc
index ef375a4d98f6d..14bdbb53b3222 100644
--- a/lib/tsan/rtl/tsan_fd.cc
+++ b/lib/tsan/rtl/tsan_fd.cc
@@ -74,13 +74,14 @@ static FdDesc *fddesc(ThreadState *thr, uptr pc, int fd) {
   uptr l1 = atomic_load(pl1, memory_order_consume);
   if (l1 == 0) {
     uptr size = kTableSizeL2 * sizeof(FdDesc);
-    void *p = internal_alloc(MBlockFD, size);
+    // We need this to reside in user memory to properly catch races on it.
+    void *p = user_alloc(thr, pc, size);
     internal_memset(p, 0, size);
     MemoryResetRange(thr, (uptr)&fddesc, (uptr)p, size);
     if (atomic_compare_exchange_strong(pl1, &l1, (uptr)p, memory_order_acq_rel))
       l1 = (uptr)p;
     else
-      internal_free(p);
+      user_free(thr, pc, p);
   }
   return &((FdDesc*)l1)[fd % kTableSizeL2];  // NOLINT
 }
@@ -150,7 +151,7 @@ void FdAcquire(ThreadState *thr, uptr pc, int fd) {
   FdDesc *d = fddesc(thr, pc, fd);
   FdSync *s = d->sync;
   DPrintf("#%d: FdAcquire(%d) -> %p\n", thr->tid, fd, s);
-  MemoryRead8Byte(thr, pc, (uptr)d);
+  MemoryRead(thr, pc, (uptr)d, kSizeLog8);
   if (s)
     Acquire(thr, pc, (uptr)s);
 }
@@ -161,20 +162,20 @@ void FdRelease(ThreadState *thr, uptr pc, int fd) {
   DPrintf("#%d: FdRelease(%d) -> %p\n", thr->tid, fd, s);
   if (s)
     Release(thr, pc, (uptr)s);
-  MemoryRead8Byte(thr, pc, (uptr)d);
+  MemoryRead(thr, pc, (uptr)d, kSizeLog8);
 }
 
 void FdAccess(ThreadState *thr, uptr pc, int fd) {
   DPrintf("#%d: FdAccess(%d)\n", thr->tid, fd);
   FdDesc *d = fddesc(thr, pc, fd);
-  MemoryRead8Byte(thr, pc, (uptr)d);
+  MemoryRead(thr, pc, (uptr)d, kSizeLog8);
 }
 
 void FdClose(ThreadState *thr, uptr pc, int fd) {
   DPrintf("#%d: FdClose(%d)\n", thr->tid, fd);
   FdDesc *d = fddesc(thr, pc, fd);
   // To catch races between fd usage and close.
-  MemoryWrite8Byte(thr, pc, (uptr)d);
+  MemoryWrite(thr, pc, (uptr)d, kSizeLog8);
   // We need to clear it, because if we do not intercept any call out there
   // that creates fd, we will hit false postives.
   MemoryResetRange(thr, pc, (uptr)d, 8);
@@ -193,7 +194,7 @@ void FdDup(ThreadState *thr, uptr pc, int oldfd, int newfd) {
   DPrintf("#%d: FdDup(%d, %d)\n", thr->tid, oldfd, newfd);
   // Ignore the case when user dups not yet connected socket.
   FdDesc *od = fddesc(thr, pc, oldfd);
-  MemoryRead8Byte(thr, pc, (uptr)od);
+  MemoryRead(thr, pc, (uptr)od, kSizeLog8);
   FdClose(thr, pc, newfd);
   init(thr, pc, newfd, ref(od->sync));
 }
diff --git a/lib/tsan/rtl/tsan_flags.cc b/lib/tsan/rtl/tsan_flags.cc
index 88c4bb6a2e441..c062592f482d8 100644
--- a/lib/tsan/rtl/tsan_flags.cc
+++ b/lib/tsan/rtl/tsan_flags.cc
@@ -45,15 +45,19 @@ void InitializeFlags(Flags *f, const char *env) {
   f->report_thread_leaks = true;
   f->report_destroy_locked = true;
   f->report_signal_unsafe = true;
+  f->report_atomic_races = true;
   f->force_seq_cst_atomics = false;
   f->strip_path_prefix = "";
   f->suppressions = "";
+  f->print_suppressions = false;
+  f->print_benign = false;
   f->exitcode = 66;
   f->log_path = "stderr";
   f->atexit_sleep_ms = 1000;
   f->verbosity = 0;
   f->profile_memory = "";
   f->flush_memory_ms = 0;
+  f->flush_symbolizer_ms = 5000;
   f->stop_on_start = false;
   f->running_on_valgrind = false;
   f->external_symbolizer_path = "";
@@ -72,15 +76,19 @@ void InitializeFlags(Flags *f, const char *env) {
   ParseFlag(env, &f->report_thread_leaks, "report_thread_leaks");
   ParseFlag(env, &f->report_destroy_locked, "report_destroy_locked");
   ParseFlag(env, &f->report_signal_unsafe, "report_signal_unsafe");
+  ParseFlag(env, &f->report_atomic_races, "report_atomic_races");
   ParseFlag(env, &f->force_seq_cst_atomics, "force_seq_cst_atomics");
   ParseFlag(env, &f->strip_path_prefix, "strip_path_prefix");
   ParseFlag(env, &f->suppressions, "suppressions");
+  ParseFlag(env, &f->print_suppressions, "print_suppressions");
+  ParseFlag(env, &f->print_benign, "print_benign");
   ParseFlag(env, &f->exitcode, "exitcode");
   ParseFlag(env, &f->log_path, "log_path");
   ParseFlag(env, &f->atexit_sleep_ms, "atexit_sleep_ms");
   ParseFlag(env, &f->verbosity, "verbosity");
   ParseFlag(env, &f->profile_memory, "profile_memory");
   ParseFlag(env, &f->flush_memory_ms, "flush_memory_ms");
+  ParseFlag(env, &f->flush_symbolizer_ms, "flush_symbolizer_ms");
   ParseFlag(env, &f->stop_on_start, "stop_on_start");
   ParseFlag(env, &f->external_symbolizer_path, "external_symbolizer_path");
   ParseFlag(env, &f->history_size, "history_size");
diff --git a/lib/tsan/rtl/tsan_flags.h b/lib/tsan/rtl/tsan_flags.h
index 6547911ec7a3b..aaacd98a6223b 100644
--- a/lib/tsan/rtl/tsan_flags.h
+++ b/lib/tsan/rtl/tsan_flags.h
@@ -43,6 +43,8 @@ struct Flags {
   // Report violations of async signal-safety
   // (e.g. malloc() call from a signal handler).
   bool report_signal_unsafe;
+  // Report races between atomic and plain memory accesses.
+  bool report_atomic_races;
   // If set, all atomics are effectively sequentially consistent (seq_cst),
   // regardless of what user actually specified.
   bool force_seq_cst_atomics;
@@ -50,6 +52,10 @@ struct Flags {
   const char *strip_path_prefix;
   // Suppressions filename.
   const char *suppressions;
+  // Print matched suppressions at exit.
+  bool print_suppressions;
+  // Print matched "benign" races at exit.
+  bool print_benign;
   // Override exit status if something was reported.
   int exitcode;
   // Write logs to "log_path.pid".
@@ -65,6 +71,8 @@ struct Flags {
   const char *profile_memory;
   // Flush shadow memory every X ms.
   int flush_memory_ms;
+  // Flush symbolizer caches every X ms.
+  int flush_symbolizer_ms;
   // Stops on start until __tsan_resume() is called (for debugging).
   bool stop_on_start;
   // Controls whether RunningOnValgrind() returns true or false.
@@ -86,6 +94,6 @@ struct Flags {
 
 Flags *flags();
 void InitializeFlags(Flags *flags, const char *env);
-}
+}  // namespace __tsan
 
 #endif  // TSAN_FLAGS_H
diff --git a/lib/tsan/rtl/tsan_interceptors.cc b/lib/tsan/rtl/tsan_interceptors.cc
index be58ca92cf915..f18b26f6abe46 100644
--- a/lib/tsan/rtl/tsan_interceptors.cc
+++ b/lib/tsan/rtl/tsan_interceptors.cc
@@ -10,11 +10,13 @@
 // This file is a part of ThreadSanitizer (TSan), a race detector.
 //
 // FIXME: move as many interceptors as possible into
-// sanitizer_common/sanitizer_common_interceptors.h
+// sanitizer_common/sanitizer_common_interceptors.inc
 //===----------------------------------------------------------------------===//
 
 #include "sanitizer_common/sanitizer_atomic.h"
 #include "sanitizer_common/sanitizer_libc.h"
+#include "sanitizer_common/sanitizer_linux.h"
+#include "sanitizer_common/sanitizer_platform_limits_posix.h"
 #include "sanitizer_common/sanitizer_placement_new.h"
 #include "sanitizer_common/sanitizer_stacktrace.h"
 #include "interception/interception.h"
@@ -26,18 +28,21 @@
 
 using namespace __tsan;  // NOLINT
 
-const int kSigCount = 128;
+const int kSigCount = 64;
 
 struct my_siginfo_t {
-  int opaque[128];
+  // The size is determined by looking at sizeof of real siginfo_t on linux.
+  u64 opaque[128 / sizeof(u64)];
 };
 
 struct sigset_t {
-  u64 val[1024 / 8 / sizeof(u64)];
+  // The size is determined by looking at sizeof of real sigset_t on linux.
+  u64 val[128 / sizeof(u64)];
 };
 
 struct ucontext_t {
-  uptr opaque[117];
+  // The size is determined by looking at sizeof of real ucontext_t on linux.
+  u64 opaque[936 / sizeof(u64) + 1];
 };
 
 extern "C" int pthread_attr_init(void *attr);
@@ -53,9 +58,13 @@ extern "C" int pthread_sigmask(int how, const sigset_t *set, sigset_t *oldset);
 extern "C" int sigfillset(sigset_t *set);
 extern "C" void *pthread_self();
 extern "C" void _exit(int status);
-extern "C" int __cxa_atexit(void (*func)(void *arg), void *arg, void *dso);
 extern "C" int *__errno_location();
 extern "C" int fileno_unlocked(void *stream);
+extern "C" void *__libc_malloc(uptr size);
+extern "C" void *__libc_calloc(uptr size, uptr n);
+extern "C" void *__libc_realloc(void *ptr, uptr size);
+extern "C" void __libc_free(void *ptr);
+extern "C" int mallopt(int param, int value);
 const int PTHREAD_MUTEX_RECURSIVE = 1;
 const int PTHREAD_MUTEX_RECURSIVE_NP = 1;
 const int kPthreadAttrSize = 56;
@@ -83,11 +92,6 @@ typedef void (*sighandler_t)(int sig);
 
 #define errno (*__errno_location())
 
-union pthread_attr_t {
-  char size[kPthreadAttrSize];
-  void *align;
-};
-
 struct sigaction_t {
   union {
     sighandler_t sa_handler;
@@ -124,7 +128,7 @@ struct SignalContext {
   int pending_signal_count;
   SignalDesc pending_signals[kSigCount];
 };
-}
+}  // namespace __tsan
 
 static SignalContext *SigCtx(ThreadState *thr) {
   SignalContext *ctx = (SignalContext*)thr->signal_ctx;
@@ -240,12 +244,15 @@ class AtExitContext {
 
   typedef void(*atexit_t)();
 
-  int atexit(ThreadState *thr, uptr pc, atexit_t f) {
+  int atexit(ThreadState *thr, uptr pc, bool is_on_exit,
+             atexit_t f, void *arg) {
     Lock l(&mtx_);
     if (pos_ == kMaxAtExit)
       return 1;
     Release(thr, pc, (uptr)this);
     stack_[pos_] = f;
+    args_[pos_] = arg;
+    is_on_exits_[pos_] = is_on_exit;
     pos_++;
     return 0;
   }
@@ -254,11 +261,15 @@ class AtExitContext {
     CHECK_EQ(thr->in_rtl, 0);
     for (;;) {
       atexit_t f = 0;
+      void *arg = 0;
+      bool is_on_exit = false;
       {
         Lock l(&mtx_);
         if (pos_) {
           pos_--;
           f = stack_[pos_];
+          arg = args_[pos_];
+          is_on_exit = is_on_exits_[pos_];
           ScopedInRtl in_rtl;
           Acquire(thr, pc, (uptr)this);
         }
@@ -267,7 +278,10 @@ class AtExitContext {
         break;
       DPrintf("#%d: executing atexit func %p\n", thr->tid, f);
       CHECK_EQ(thr->in_rtl, 0);
-      f();
+      if (is_on_exit)
+        ((void(*)(int status, void *arg))f)(0, arg);
+      else
+        ((void(*)(void *arg, void *dso))f)(arg, 0);
     }
   }
 
@@ -275,42 +289,133 @@ class AtExitContext {
   static const int kMaxAtExit = 128;
   Mutex mtx_;
   atexit_t stack_[kMaxAtExit];
+  void *args_[kMaxAtExit];
+  bool is_on_exits_[kMaxAtExit];
   int pos_;
 };
 
 static AtExitContext *atexit_ctx;
 
-static void finalize(void *arg) {
-  ThreadState * thr = cur_thread();
-  uptr pc = 0;
-  atexit_ctx->exit(thr, pc);
-  {
-    ScopedInRtl in_rtl;
-    DestroyAndFree(atexit_ctx);
+TSAN_INTERCEPTOR(int, atexit, void (*f)()) {
+  if (cur_thread()->in_symbolizer)
+    return 0;
+  SCOPED_TSAN_INTERCEPTOR(atexit, f);
+  return atexit_ctx->atexit(thr, pc, false, (void(*)())f, 0);
+}
+
+TSAN_INTERCEPTOR(int, on_exit, void(*f)(int, void*), void *arg) {
+  if (cur_thread()->in_symbolizer)
+    return 0;
+  SCOPED_TSAN_INTERCEPTOR(on_exit, f, arg);
+  return atexit_ctx->atexit(thr, pc, true, (void(*)())f, arg);
+}
+
+TSAN_INTERCEPTOR(int, __cxa_atexit, void (*f)(void *a), void *arg, void *dso) {
+  if (cur_thread()->in_symbolizer)
+    return 0;
+  SCOPED_TSAN_INTERCEPTOR(__cxa_atexit, f, arg, dso);
+  if (dso)
+    return REAL(__cxa_atexit)(f, arg, dso);
+  return atexit_ctx->atexit(thr, pc, false, (void(*)())f, arg);
+}
+
+// Cleanup old bufs.
+static void JmpBufGarbageCollect(ThreadState *thr, uptr sp) {
+  for (uptr i = 0; i < thr->jmp_bufs.Size(); i++) {
+    JmpBuf *buf = &thr->jmp_bufs[i];
+    if (buf->sp <= sp) {
+      uptr sz = thr->jmp_bufs.Size();
+      thr->jmp_bufs[i] = thr->jmp_bufs[sz - 1];
+      thr->jmp_bufs.PopBack();
+      i--;
+    }
   }
-  int status = Finalize(cur_thread());
-  if (status)
-    _exit(status);
 }
 
-TSAN_INTERCEPTOR(int, atexit, void (*f)()) {
-  SCOPED_TSAN_INTERCEPTOR(atexit, f);
-  return atexit_ctx->atexit(thr, pc, f);
+static void SetJmp(ThreadState *thr, uptr sp, uptr mangled_sp) {
+  if (thr->shadow_stack_pos == 0)  // called from libc guts during bootstrap
+    return;
+  // Cleanup old bufs.
+  JmpBufGarbageCollect(thr, sp);
+  // Remember the buf.
+  JmpBuf *buf = thr->jmp_bufs.PushBack();
+  buf->sp = sp;
+  buf->mangled_sp = mangled_sp;
+  buf->shadow_stack_pos = thr->shadow_stack_pos;
+}
+
+static void LongJmp(ThreadState *thr, uptr *env) {
+  uptr mangled_sp = env[6];
+  // Find the saved buf by mangled_sp.
+  for (uptr i = 0; i < thr->jmp_bufs.Size(); i++) {
+    JmpBuf *buf = &thr->jmp_bufs[i];
+    if (buf->mangled_sp == mangled_sp) {
+      CHECK_GE(thr->shadow_stack_pos, buf->shadow_stack_pos);
+      // Unwind the stack.
+      while (thr->shadow_stack_pos > buf->shadow_stack_pos)
+        FuncExit(thr);
+      JmpBufGarbageCollect(thr, buf->sp - 1);  // do not collect buf->sp
+      return;
+    }
+  }
+  Printf("ThreadSanitizer: can't find longjmp buf\n");
+  CHECK(0);
 }
 
-TSAN_INTERCEPTOR(void, longjmp, void *env, int val) {
-  SCOPED_TSAN_INTERCEPTOR(longjmp, env, val);
-  Printf("ThreadSanitizer: longjmp() is not supported\n");
-  Die();
+extern "C" void __tsan_setjmp(uptr sp, uptr mangled_sp) {
+  ScopedInRtl in_rtl;
+  SetJmp(cur_thread(), sp, mangled_sp);
 }
 
-TSAN_INTERCEPTOR(void, siglongjmp, void *env, int val) {
-  SCOPED_TSAN_INTERCEPTOR(siglongjmp, env, val);
-  Printf("ThreadSanitizer: siglongjmp() is not supported\n");
-  Die();
+// Not called.  Merely to satisfy TSAN_INTERCEPT().
+extern "C" int __interceptor_setjmp(void *env) {
+  CHECK(0);
+  return 0;
+}
+
+extern "C" int __interceptor__setjmp(void *env) {
+  CHECK(0);
+  return 0;
+}
+
+extern "C" int __interceptor_sigsetjmp(void *env) {
+  CHECK(0);
+  return 0;
+}
+
+extern "C" int __interceptor___sigsetjmp(void *env) {
+  CHECK(0);
+  return 0;
+}
+
+extern "C" int setjmp(void *env);
+extern "C" int _setjmp(void *env);
+extern "C" int sigsetjmp(void *env);
+extern "C" int __sigsetjmp(void *env);
+DEFINE_REAL(int, setjmp, void *env)
+DEFINE_REAL(int, _setjmp, void *env)
+DEFINE_REAL(int, sigsetjmp, void *env)
+DEFINE_REAL(int, __sigsetjmp, void *env)
+
+TSAN_INTERCEPTOR(void, longjmp, uptr *env, int val) {
+  {
+    SCOPED_TSAN_INTERCEPTOR(longjmp, env, val);
+  }
+  LongJmp(cur_thread(), env);
+  REAL(longjmp)(env, val);
+}
+
+TSAN_INTERCEPTOR(void, siglongjmp, uptr *env, int val) {
+  {
+    SCOPED_TSAN_INTERCEPTOR(siglongjmp, env, val);
+  }
+  LongJmp(cur_thread(), env);
+  REAL(siglongjmp)(env, val);
 }
 
 TSAN_INTERCEPTOR(void*, malloc, uptr size) {
+  if (cur_thread()->in_symbolizer)
+    return __libc_malloc(size);
   void *p = 0;
   {
     SCOPED_INTERCEPTOR_RAW(malloc, size);
@@ -326,17 +431,23 @@ TSAN_INTERCEPTOR(void*, __libc_memalign, uptr align, uptr sz) {
 }
 
 TSAN_INTERCEPTOR(void*, calloc, uptr size, uptr n) {
+  if (cur_thread()->in_symbolizer)
+    return __libc_calloc(size, n);
+  if (__sanitizer::CallocShouldReturnNullDueToOverflow(size, n)) return 0;
   void *p = 0;
   {
     SCOPED_INTERCEPTOR_RAW(calloc, size, n);
     p = user_alloc(thr, pc, n * size);
-    if (p) internal_memset(p, 0, n * size);
+    if (p)
+      internal_memset(p, 0, n * size);
   }
   invoke_malloc_hook(p, n * size);
   return p;
 }
 
 TSAN_INTERCEPTOR(void*, realloc, void *p, uptr size) {
+  if (cur_thread()->in_symbolizer)
+    return __libc_realloc(p, size);
   if (p)
     invoke_free_hook(p);
   {
@@ -350,6 +461,8 @@ TSAN_INTERCEPTOR(void*, realloc, void *p, uptr size) {
 TSAN_INTERCEPTOR(void, free, void *p) {
   if (p == 0)
     return;
+  if (cur_thread()->in_symbolizer)
+    return __libc_free(p);
   invoke_free_hook(p);
   SCOPED_INTERCEPTOR_RAW(free, p);
   user_free(thr, pc, p);
@@ -358,12 +471,21 @@ TSAN_INTERCEPTOR(void, free, void *p) {
 TSAN_INTERCEPTOR(void, cfree, void *p) {
   if (p == 0)
     return;
+  if (cur_thread()->in_symbolizer)
+    return __libc_free(p);
   invoke_free_hook(p);
   SCOPED_INTERCEPTOR_RAW(cfree, p);
   user_free(thr, pc, p);
 }
 
+TSAN_INTERCEPTOR(uptr, malloc_usable_size, void *p) {
+  SCOPED_INTERCEPTOR_RAW(malloc_usable_size, p);
+  return user_alloc_usable_size(thr, pc, p);
+}
+
 #define OPERATOR_NEW_BODY(mangled_name) \
+  if (cur_thread()->in_symbolizer) \
+    return __libc_malloc(size); \
   void *p = 0; \
   {  \
     SCOPED_INTERCEPTOR_RAW(mangled_name, size); \
@@ -387,6 +509,8 @@ void *operator new[](__sanitizer::uptr size, std::nothrow_t const&) {
 
 #define OPERATOR_DELETE_BODY(mangled_name) \
   if (ptr == 0) return;  \
+  if (cur_thread()->in_symbolizer) \
+    return __libc_free(ptr); \
   invoke_free_hook(ptr);  \
   SCOPED_INTERCEPTOR_RAW(mangled_name, ptr);  \
   user_free(thr, pc, ptr);
@@ -551,7 +675,9 @@ TSAN_INTERCEPTOR(void*, mmap, void *addr, long_t sz, int prot,
     return MAP_FAILED;
   void *res = REAL(mmap)(addr, sz, prot, flags, fd, off);
   if (res != MAP_FAILED) {
-    MemoryResetRange(thr, pc, (uptr)res, sz);
+    if (fd > 0)
+      FdAccess(thr, pc, fd);
+    MemoryRangeImitateWrite(thr, pc, (uptr)res, sz);
   }
   return res;
 }
@@ -563,13 +689,16 @@ TSAN_INTERCEPTOR(void*, mmap64, void *addr, long_t sz, int prot,
     return MAP_FAILED;
   void *res = REAL(mmap64)(addr, sz, prot, flags, fd, off);
   if (res != MAP_FAILED) {
-    MemoryResetRange(thr, pc, (uptr)res, sz);
+    if (fd > 0)
+      FdAccess(thr, pc, fd);
+    MemoryRangeImitateWrite(thr, pc, (uptr)res, sz);
   }
   return res;
 }
 
 TSAN_INTERCEPTOR(int, munmap, void *addr, long_t sz) {
   SCOPED_TSAN_INTERCEPTOR(munmap, addr, sz);
+  DontNeedShadowFor((uptr)addr, sz);
   int res = REAL(munmap)(addr, sz);
   return res;
 }
@@ -681,21 +810,21 @@ extern "C" void *__tsan_thread_start_func(void *arg) {
 TSAN_INTERCEPTOR(int, pthread_create,
     void *th, void *attr, void *(*callback)(void*), void * param) {
   SCOPED_TSAN_INTERCEPTOR(pthread_create, th, attr, callback, param);
-  pthread_attr_t myattr;
+  __sanitizer_pthread_attr_t myattr;
   if (attr == 0) {
     pthread_attr_init(&myattr);
     attr = &myattr;
   }
   int detached = 0;
   pthread_attr_getdetachstate(attr, &detached);
-  uptr stacksize = 0;
-  pthread_attr_getstacksize(attr, &stacksize);
-  // We place the huge ThreadState object into TLS, account for that.
-  const uptr minstacksize = GetTlsSize() + 128*1024;
-  if (stacksize < minstacksize) {
-    DPrintf("ThreadSanitizer: stacksize %zu->%zu\n", stacksize, minstacksize);
-    pthread_attr_setstacksize(attr, minstacksize);
-  }
+
+#if defined(TSAN_DEBUG_OUTPUT)
+  int verbosity = (TSAN_DEBUG_OUTPUT);
+#else
+  int verbosity = 0;
+#endif
+  AdjustStackSizeLinux(attr, verbosity);
+
   ThreadParam p;
   p.callback = callback;
   p.param = param;
@@ -960,14 +1089,14 @@ TSAN_INTERCEPTOR(int, pthread_cond_timedwait, void *c, void *m, void *abstime) {
 
 TSAN_INTERCEPTOR(int, pthread_barrier_init, void *b, void *a, unsigned count) {
   SCOPED_TSAN_INTERCEPTOR(pthread_barrier_init, b, a, count);
-  MemoryWrite1Byte(thr, pc, (uptr)b);
+  MemoryWrite(thr, pc, (uptr)b, kSizeLog1);
   int res = REAL(pthread_barrier_init)(b, a, count);
   return res;
 }
 
 TSAN_INTERCEPTOR(int, pthread_barrier_destroy, void *b) {
   SCOPED_TSAN_INTERCEPTOR(pthread_barrier_destroy, b);
-  MemoryWrite1Byte(thr, pc, (uptr)b);
+  MemoryWrite(thr, pc, (uptr)b, kSizeLog1);
   int res = REAL(pthread_barrier_destroy)(b);
   return res;
 }
@@ -975,9 +1104,9 @@ TSAN_INTERCEPTOR(int, pthread_barrier_destroy, void *b) {
 TSAN_INTERCEPTOR(int, pthread_barrier_wait, void *b) {
   SCOPED_TSAN_INTERCEPTOR(pthread_barrier_wait, b);
   Release(thr, pc, (uptr)b);
-  MemoryRead1Byte(thr, pc, (uptr)b);
+  MemoryRead(thr, pc, (uptr)b, kSizeLog1);
   int res = REAL(pthread_barrier_wait)(b);
-  MemoryRead1Byte(thr, pc, (uptr)b);
+  MemoryRead(thr, pc, (uptr)b, kSizeLog1);
   if (res == 0 || res == PTHREAD_BARRIER_SERIAL_THREAD) {
     Acquire(thr, pc, (uptr)b);
   }
@@ -1064,6 +1193,74 @@ TSAN_INTERCEPTOR(int, sem_getvalue, void *s, int *sval) {
   return res;
 }
 
+TSAN_INTERCEPTOR(int, __xstat, int version, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__xstat, version, path, buf);
+  return REAL(__xstat)(version, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, stat, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__xstat, 0, path, buf);
+  return REAL(__xstat)(0, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, __xstat64, int version, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__xstat64, version, path, buf);
+  return REAL(__xstat64)(version, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, stat64, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__xstat64, 0, path, buf);
+  return REAL(__xstat64)(0, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, __lxstat, int version, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__lxstat, version, path, buf);
+  return REAL(__lxstat)(version, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, lstat, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__lxstat, 0, path, buf);
+  return REAL(__lxstat)(0, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, __lxstat64, int version, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__lxstat64, version, path, buf);
+  return REAL(__lxstat64)(version, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, lstat64, const char *path, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__lxstat64, 0, path, buf);
+  return REAL(__lxstat64)(0, path, buf);
+}
+
+TSAN_INTERCEPTOR(int, __fxstat, int version, int fd, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__fxstat, version, fd, buf);
+  if (fd > 0)
+    FdAccess(thr, pc, fd);
+  return REAL(__fxstat)(version, fd, buf);
+}
+
+TSAN_INTERCEPTOR(int, fstat, int fd, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__fxstat, 0, fd, buf);
+  if (fd > 0)
+    FdAccess(thr, pc, fd);
+  return REAL(__fxstat)(0, fd, buf);
+}
+
+TSAN_INTERCEPTOR(int, __fxstat64, int version, int fd, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__fxstat64, version, fd, buf);
+  if (fd > 0)
+    FdAccess(thr, pc, fd);
+  return REAL(__fxstat64)(version, fd, buf);
+}
+
+TSAN_INTERCEPTOR(int, fstat64, int fd, void *buf) {
+  SCOPED_TSAN_INTERCEPTOR(__fxstat64, 0, fd, buf);
+  if (fd > 0)
+    FdAccess(thr, pc, fd);
+  return REAL(__fxstat64)(0, fd, buf);
+}
+
 TSAN_INTERCEPTOR(int, open, const char *name, int flags, int mode) {
   SCOPED_TSAN_INTERCEPTOR(open, name, flags, mode);
   int fd = REAL(open)(name, flags, mode);
@@ -1179,6 +1376,22 @@ TSAN_INTERCEPTOR(int, connect, int fd, void *addr, unsigned addrlen) {
   return res;
 }
 
+TSAN_INTERCEPTOR(int, bind, int fd, void *addr, unsigned addrlen) {
+  SCOPED_TSAN_INTERCEPTOR(bind, fd, addr, addrlen);
+  int res = REAL(bind)(fd, addr, addrlen);
+  if (fd > 0 && res == 0)
+    FdAccess(thr, pc, fd);
+  return res;
+}
+
+TSAN_INTERCEPTOR(int, listen, int fd, int backlog) {
+  SCOPED_TSAN_INTERCEPTOR(listen, fd, backlog);
+  int res = REAL(listen)(fd, backlog);
+  if (fd > 0 && res == 0)
+    FdAccess(thr, pc, fd);
+  return res;
+}
+
 TSAN_INTERCEPTOR(int, accept, int fd, void *addr, unsigned *addrlen) {
   SCOPED_TSAN_INTERCEPTOR(accept, fd, addr, addrlen);
   int fd2 = REAL(accept)(fd, addr, addrlen);
@@ -1225,6 +1438,18 @@ TSAN_INTERCEPTOR(int, __close, int fd) {
   return REAL(__close)(fd);
 }
 
+// glibc guts
+TSAN_INTERCEPTOR(void, __res_iclose, void *state, bool free_addr) {
+  SCOPED_TSAN_INTERCEPTOR(__res_iclose, state, free_addr);
+  int fds[64];
+  int cnt = ExtractResolvFDs(state, fds, ARRAY_SIZE(fds));
+  for (int i = 0; i < cnt; i++) {
+    if (fds[i] > 0)
+      FdClose(thr, pc, fds[i]);
+  }
+  REAL(__res_iclose)(state, free_addr);
+}
+
 TSAN_INTERCEPTOR(int, pipe, int *pipefd) {
   SCOPED_TSAN_INTERCEPTOR(pipe, pipefd);
   int res = REAL(pipe)(pipefd);
@@ -1373,6 +1598,17 @@ TSAN_INTERCEPTOR(uptr, fwrite, const void *p, uptr size, uptr nmemb, void *f) {
   return REAL(fwrite)(p, size, nmemb, f);
 }
 
+TSAN_INTERCEPTOR(int, fflush, void *stream) {
+  SCOPED_TSAN_INTERCEPTOR(fflush, stream);
+  return REAL(fflush)(stream);
+}
+
+TSAN_INTERCEPTOR(void, abort, int fake) {
+  SCOPED_TSAN_INTERCEPTOR(abort, fake);
+  REAL(fflush)(0);
+  REAL(abort)(fake);
+}
+
 TSAN_INTERCEPTOR(int, puts, const char *s) {
   SCOPED_TSAN_INTERCEPTOR(puts, s);
   MemoryAccessRange(thr, pc, (uptr)s, internal_strlen(s), false);
@@ -1420,7 +1656,7 @@ TSAN_INTERCEPTOR(int, poll, void *fds, long_t nfds, int timeout) {
   return res;
 }
 
-static void ALWAYS_INLINE rtl_generic_sighandler(bool sigact, int sig,
+void ALWAYS_INLINE rtl_generic_sighandler(bool sigact, int sig,
     my_siginfo_t *info, void *ctx) {
   ThreadState *thr = cur_thread();
   SignalContext *sctx = SigCtx(thr);
@@ -1433,7 +1669,6 @@ static void ALWAYS_INLINE rtl_generic_sighandler(bool sigact, int sig,
       // (but check if we are in a recursive interceptor,
       // i.e. pthread_join()->munmap()).
       (sctx && sctx->in_blocking_func == 1 && thr->in_rtl == 1)) {
-    CHECK(thr->in_rtl == 0 || thr->in_rtl == 1);
     int in_rtl = thr->in_rtl;
     thr->in_rtl = 0;
     CHECK_EQ(thr->in_signal_handler, false);
@@ -1519,11 +1754,11 @@ TSAN_INTERCEPTOR(int, kill, int pid, int sig) {
   SignalContext *sctx = SigCtx(thr);
   CHECK_NE(sctx, 0);
   int prev = sctx->int_signal_send;
-  if (pid == GetPid()) {
+  if (pid == (int)internal_getpid()) {
     sctx->int_signal_send = sig;
   }
   int res = REAL(kill)(pid, sig);
-  if (pid == GetPid()) {
+  if (pid == (int)internal_getpid()) {
     CHECK_EQ(sctx->int_signal_send, sig);
     sctx->int_signal_send = prev;
   }
@@ -1600,6 +1835,13 @@ struct TsanInterceptorContext {
   const uptr pc;
 };
 
+#include "sanitizer_common/sanitizer_platform_interceptors.h"
+// Causes interceptor recursion (getpwuid_r() calls fopen())
+#undef SANITIZER_INTERCEPT_GETPWNAM_AND_FRIENDS
+#undef SANITIZER_INTERCEPT_GETPWNAM_R_AND_FRIENDS
+// Causes interceptor recursion (glob64() calls lstat64())
+#undef SANITIZER_INTERCEPT_GLOB
+
 #define COMMON_INTERCEPTOR_WRITE_RANGE(ctx, ptr, size) \
     MemoryAccessRange(((TsanInterceptorContext*)ctx)->thr,  \
                       ((TsanInterceptorContext*)ctx)->pc,   \
@@ -1621,6 +1863,13 @@ struct TsanInterceptorContext {
     ThreadSetName(((TsanInterceptorContext*)ctx)->thr, name)
 #include "sanitizer_common/sanitizer_common_interceptors.inc"
 
+// FIXME: Implement these with MemoryAccessRange().
+#define COMMON_SYSCALL_PRE_READ_RANGE(p, s)
+#define COMMON_SYSCALL_PRE_WRITE_RANGE(p, s)
+#define COMMON_SYSCALL_POST_READ_RANGE(p, s)
+#define COMMON_SYSCALL_POST_WRITE_RANGE(p, s)
+#include "sanitizer_common/sanitizer_common_syscalls.inc"
+
 namespace __tsan {
 
 void ProcessPendingSignals(ThreadState *thr) {
@@ -1655,7 +1904,7 @@ void ProcessPendingSignals(ThreadState *thr) {
               (uptr)sigactions[sig].sa_sigaction :
               (uptr)sigactions[sig].sa_handler;
           stack.Init(&pc, 1);
-          Lock l(&ctx->thread_mtx);
+          ThreadRegistryLock l(ctx->thread_registry);
           ScopedReport rep(ReportTypeErrnoInSignal);
           if (!IsFiredSuppression(ctx, rep, stack)) {
             rep.AddStack(&stack);
@@ -1671,6 +1920,16 @@ void ProcessPendingSignals(ThreadState *thr) {
   thr->in_signal_handler = false;
 }
 
+static void finalize(void *arg) {
+  ThreadState * thr = cur_thread();
+  uptr pc = 0;
+  atexit_ctx->exit(thr, pc);
+  int status = Finalize(cur_thread());
+  REAL(fflush)(0);
+  if (status)
+    _exit(status);
+}
+
 static void unreachable() {
   Printf("FATAL: ThreadSanitizer: unreachable called\n");
   Die();
@@ -1684,8 +1943,16 @@ void InitializeInterceptors() {
   REAL(memcpy) = internal_memcpy;
   REAL(memcmp) = internal_memcmp;
 
+  // Instruct libc malloc to consume less memory.
+  mallopt(1, 0);  // M_MXFAST
+  mallopt(-3, 32*1024);  // M_MMAP_THRESHOLD
+
   SANITIZER_COMMON_INTERCEPTORS_INIT;
 
+  TSAN_INTERCEPT(setjmp);
+  TSAN_INTERCEPT(_setjmp);
+  TSAN_INTERCEPT(sigsetjmp);
+  TSAN_INTERCEPT(__sigsetjmp);
   TSAN_INTERCEPT(longjmp);
   TSAN_INTERCEPT(siglongjmp);
 
@@ -1767,6 +2034,18 @@ void InitializeInterceptors() {
   TSAN_INTERCEPT(sem_post);
   TSAN_INTERCEPT(sem_getvalue);
 
+  TSAN_INTERCEPT(stat);
+  TSAN_INTERCEPT(__xstat);
+  TSAN_INTERCEPT(stat64);
+  TSAN_INTERCEPT(__xstat64);
+  TSAN_INTERCEPT(lstat);
+  TSAN_INTERCEPT(__lxstat);
+  TSAN_INTERCEPT(lstat64);
+  TSAN_INTERCEPT(__lxstat64);
+  TSAN_INTERCEPT(fstat);
+  TSAN_INTERCEPT(__fxstat);
+  TSAN_INTERCEPT(fstat64);
+  TSAN_INTERCEPT(__fxstat64);
   TSAN_INTERCEPT(open);
   TSAN_INTERCEPT(open64);
   TSAN_INTERCEPT(creat);
@@ -1781,11 +2060,15 @@ void InitializeInterceptors() {
   TSAN_INTERCEPT(socket);
   TSAN_INTERCEPT(socketpair);
   TSAN_INTERCEPT(connect);
+  TSAN_INTERCEPT(bind);
+  TSAN_INTERCEPT(listen);
   TSAN_INTERCEPT(accept);
   TSAN_INTERCEPT(accept4);
   TSAN_INTERCEPT(epoll_create);
   TSAN_INTERCEPT(epoll_create1);
   TSAN_INTERCEPT(close);
+  TSAN_INTERCEPT(__close);
+  TSAN_INTERCEPT(__res_iclose);
   TSAN_INTERCEPT(pipe);
   TSAN_INTERCEPT(pipe2);
 
@@ -1804,6 +2087,8 @@ void InitializeInterceptors() {
   TSAN_INTERCEPT(fclose);
   TSAN_INTERCEPT(fread);
   TSAN_INTERCEPT(fwrite);
+  TSAN_INTERCEPT(fflush);
+  TSAN_INTERCEPT(abort);
   TSAN_INTERCEPT(puts);
   TSAN_INTERCEPT(rmdir);
   TSAN_INTERCEPT(opendir);
@@ -1828,6 +2113,8 @@ void InitializeInterceptors() {
   TSAN_INTERCEPT(munlockall);
 
   TSAN_INTERCEPT(fork);
+  TSAN_INTERCEPT(on_exit);
+  TSAN_INTERCEPT(__cxa_atexit);
 
   // Need to setup it, because interceptors check that the function is resolved.
   // But atexit is emitted directly into the module, so can't be resolved.
@@ -1835,7 +2122,7 @@ void InitializeInterceptors() {
   atexit_ctx = new(internal_alloc(MBlockAtExit, sizeof(AtExitContext)))
       AtExitContext();
 
-  if (__cxa_atexit(&finalize, 0, 0)) {
+  if (REAL(__cxa_atexit)(&finalize, 0, 0)) {
     Printf("ThreadSanitizer: failed to setup atexit callback\n");
     Die();
   }
diff --git a/lib/tsan/rtl/tsan_interface.cc b/lib/tsan/rtl/tsan_interface.cc
index 6d0954602ff74..efad8c192d6e5 100644
--- a/lib/tsan/rtl/tsan_interface.cc
+++ b/lib/tsan/rtl/tsan_interface.cc
@@ -14,23 +14,73 @@
 #include "tsan_interface.h"
 #include "tsan_interface_ann.h"
 #include "tsan_rtl.h"
+#include "sanitizer_common/sanitizer_internal_defs.h"
 
 #define CALLERPC ((uptr)__builtin_return_address(0))
 
 using namespace __tsan;  // NOLINT
 
+typedef u16 uint16_t;
+typedef u32 uint32_t;
+typedef u64 uint64_t;
+
 void __tsan_init() {
   Initialize(cur_thread());
 }
 
 void __tsan_read16(void *addr) {
-  MemoryRead8Byte(cur_thread(), CALLERPC, (uptr)addr);
-  MemoryRead8Byte(cur_thread(), CALLERPC, (uptr)addr + 8);
+  MemoryRead(cur_thread(), CALLERPC, (uptr)addr, kSizeLog8);
+  MemoryRead(cur_thread(), CALLERPC, (uptr)addr + 8, kSizeLog8);
 }
 
 void __tsan_write16(void *addr) {
-  MemoryWrite8Byte(cur_thread(), CALLERPC, (uptr)addr);
-  MemoryWrite8Byte(cur_thread(), CALLERPC, (uptr)addr + 8);
+  MemoryWrite(cur_thread(), CALLERPC, (uptr)addr, kSizeLog8);
+  MemoryWrite(cur_thread(), CALLERPC, (uptr)addr + 8, kSizeLog8);
+}
+
+u16 __tsan_unaligned_read2(void *addr) {
+  UnalignedMemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 2, false, false);
+  return *(u16*)addr;
+}
+
+u32 __tsan_unaligned_read4(void *addr) {
+  UnalignedMemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 4, false, false);
+  return *(u32*)addr;
+}
+
+u64 __tsan_unaligned_read8(void *addr) {
+  UnalignedMemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 8, false, false);
+  return *(u64*)addr;
+}
+
+void __tsan_unaligned_write2(void *addr, u16 v) {
+  UnalignedMemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 2, true, false);
+  *(u16*)addr = v;
+}
+
+void __tsan_unaligned_write4(void *addr, u32 v) {
+  UnalignedMemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 4, true, false);
+  *(u32*)addr = v;
+}
+
+void __tsan_unaligned_write8(void *addr, u64 v) {
+  UnalignedMemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 8, true, false);
+  *(u64*)addr = v;
+}
+
+extern "C" {
+uint16_t __sanitizer_unaligned_load16(void *addr)
+    ALIAS("__tsan_unaligned_read2") SANITIZER_INTERFACE_ATTRIBUTE;
+uint32_t __sanitizer_unaligned_load32(void *addr)
+    ALIAS("__tsan_unaligned_read4") SANITIZER_INTERFACE_ATTRIBUTE;
+uint64_t __sanitizer_unaligned_load64(void *addr)
+    ALIAS("__tsan_unaligned_read8") SANITIZER_INTERFACE_ATTRIBUTE;
+void __sanitizer_unaligned_store16(void *addr, uint16_t v)
+    ALIAS("__tsan_unaligned_write2") SANITIZER_INTERFACE_ATTRIBUTE;
+void __sanitizer_unaligned_store32(void *addr, uint32_t v)
+    ALIAS("__tsan_unaligned_write4") SANITIZER_INTERFACE_ATTRIBUTE;
+void __sanitizer_unaligned_store64(void *addr, uint64_t v)
+    ALIAS("__tsan_unaligned_write8") SANITIZER_INTERFACE_ATTRIBUTE;
 }
 
 void __tsan_acquire(void *addr) {
diff --git a/lib/tsan/rtl/tsan_interface.h b/lib/tsan/rtl/tsan_interface.h
index 7480fc893f2d0..457fb55e0d2d1 100644
--- a/lib/tsan/rtl/tsan_interface.h
+++ b/lib/tsan/rtl/tsan_interface.h
@@ -16,7 +16,7 @@
 #ifndef TSAN_INTERFACE_H
 #define TSAN_INTERFACE_H
 
-#include <sanitizer/common_interface_defs.h>
+#include <sanitizer_common/sanitizer_internal_defs.h>
 
 // This header should NOT include any other headers.
 // All functions in this header are extern "C" and start with __tsan_.
@@ -41,6 +41,14 @@ void __tsan_write4(void *addr) SANITIZER_INTERFACE_ATTRIBUTE;
 void __tsan_write8(void *addr) SANITIZER_INTERFACE_ATTRIBUTE;
 void __tsan_write16(void *addr) SANITIZER_INTERFACE_ATTRIBUTE;
 
+u16 __tsan_unaligned_read2(void *addr) SANITIZER_INTERFACE_ATTRIBUTE;
+u32 __tsan_unaligned_read4(void *addr) SANITIZER_INTERFACE_ATTRIBUTE;
+u64 __tsan_unaligned_read8(void *addr) SANITIZER_INTERFACE_ATTRIBUTE;
+void __tsan_unaligned_write2(void *addr, u16 v) SANITIZER_INTERFACE_ATTRIBUTE;
+void __tsan_unaligned_write4(void *addr, u32 v) SANITIZER_INTERFACE_ATTRIBUTE;
+void __tsan_unaligned_write8(void *addr, u64 v) SANITIZER_INTERFACE_ATTRIBUTE;
+
+void __tsan_vptr_read(void **vptr_p) SANITIZER_INTERFACE_ATTRIBUTE;
 void __tsan_vptr_update(void **vptr_p, void *new_val)
     SANITIZER_INTERFACE_ATTRIBUTE;
 
diff --git a/lib/tsan/rtl/tsan_interface_ann.cc b/lib/tsan/rtl/tsan_interface_ann.cc
index 51ebbf2266dd9..04b4b455d15ef 100644
--- a/lib/tsan/rtl/tsan_interface_ann.cc
+++ b/lib/tsan/rtl/tsan_interface_ann.cc
@@ -20,6 +20,7 @@
 #include "tsan_mman.h"
 #include "tsan_flags.h"
 #include "tsan_platform.h"
+#include "tsan_vector.h"
 
 #define CALLERPC ((uptr)__builtin_return_address(0))
 
@@ -67,6 +68,7 @@ struct ExpectRace {
   ExpectRace *next;
   ExpectRace *prev;
   int hitcount;
+  int addcount;
   uptr addr;
   uptr size;
   char *file;
@@ -91,16 +93,19 @@ static void AddExpectRace(ExpectRace *list,
     char *f, int l, uptr addr, uptr size, char *desc) {
   ExpectRace *race = list->next;
   for (; race != list; race = race->next) {
-    if (race->addr == addr && race->size == size)
+    if (race->addr == addr && race->size == size) {
+      race->addcount++;
       return;
+    }
   }
   race = (ExpectRace*)internal_alloc(MBlockExpectRace, sizeof(ExpectRace));
-  race->hitcount = 0;
   race->addr = addr;
   race->size = size;
   race->file = f;
   race->line = l;
   race->desc[0] = 0;
+  race->hitcount = 0;
+  race->addcount = 1;
   if (desc) {
     int i = 0;
     for (; i < kMaxDescLen - 1 && desc[i]; i++)
@@ -155,6 +160,68 @@ bool IsExpectedReport(uptr addr, uptr size) {
   return false;
 }
 
+static void CollectMatchedBenignRaces(Vector<ExpectRace> *matched,
+    int *unique_count, int *hit_count, int ExpectRace::*counter) {
+  ExpectRace *list = &dyn_ann_ctx->benign;
+  for (ExpectRace *race = list->next; race != list; race = race->next) {
+    (*unique_count)++;
+    if (race->*counter == 0)
+      continue;
+    (*hit_count) += race->*counter;
+    uptr i = 0;
+    for (; i < matched->Size(); i++) {
+      ExpectRace *race0 = &(*matched)[i];
+      if (race->line == race0->line
+          && internal_strcmp(race->file, race0->file) == 0
+          && internal_strcmp(race->desc, race0->desc) == 0) {
+        race0->*counter += race->*counter;
+        break;
+      }
+    }
+    if (i == matched->Size())
+      matched->PushBack(*race);
+  }
+}
+
+void PrintMatchedBenignRaces() {
+  Lock lock(&dyn_ann_ctx->mtx);
+  int unique_count = 0;
+  int hit_count = 0;
+  int add_count = 0;
+  Vector<ExpectRace> hit_matched(MBlockScopedBuf);
+  CollectMatchedBenignRaces(&hit_matched, &unique_count, &hit_count,
+      &ExpectRace::hitcount);
+  Vector<ExpectRace> add_matched(MBlockScopedBuf);
+  CollectMatchedBenignRaces(&add_matched, &unique_count, &add_count,
+      &ExpectRace::addcount);
+  if (hit_matched.Size()) {
+    Printf("ThreadSanitizer: Matched %d \"benign\" races (pid=%d):\n",
+        hit_count, (int)internal_getpid());
+    for (uptr i = 0; i < hit_matched.Size(); i++) {
+      Printf("%d %s:%d %s\n",
+          hit_matched[i].hitcount, hit_matched[i].file,
+          hit_matched[i].line, hit_matched[i].desc);
+    }
+  }
+  if (hit_matched.Size()) {
+    Printf("ThreadSanitizer: Annotated %d \"benign\" races, %d unique"
+           " (pid=%d):\n",
+        add_count, unique_count, (int)internal_getpid());
+    for (uptr i = 0; i < add_matched.Size(); i++) {
+      Printf("%d %s:%d %s\n",
+          add_matched[i].addcount, add_matched[i].file,
+          add_matched[i].line, add_matched[i].desc);
+    }
+  }
+}
+
+static void ReportMissedExpectedRace(ExpectRace *race) {
+  Printf("==================\n");
+  Printf("WARNING: ThreadSanitizer: missed expected data race\n");
+  Printf("  %s addr=%zx %s:%d\n",
+      race->desc, race->addr, race->file, race->line);
+  Printf("==================\n");
+}
 }  // namespace __tsan
 
 using namespace __tsan;  // NOLINT
@@ -237,14 +304,6 @@ void INTERFACE_ATTRIBUTE AnnotateNoOp(char *f, int l, uptr mem) {
   SCOPED_ANNOTATION(AnnotateNoOp);
 }
 
-static void ReportMissedExpectedRace(ExpectRace *race) {
-  Printf("==================\n");
-  Printf("WARNING: ThreadSanitizer: missed expected data race\n");
-  Printf("  %s addr=%zx %s:%d\n",
-      race->desc, race->addr, race->file, race->line);
-  Printf("==================\n");
-}
-
 void INTERFACE_ATTRIBUTE AnnotateFlushExpectedRaces(char *f, int l) {
   SCOPED_ANNOTATION(AnnotateFlushExpectedRaces);
   Lock lock(&dyn_ann_ctx->mtx);
@@ -357,6 +416,9 @@ void INTERFACE_ATTRIBUTE AnnotateThreadName(
   ThreadSetName(thr, name);
 }
 
+// We deliberately omit the implementation of WTFAnnotateHappensBefore() and
+// WTFAnnotateHappensAfter(). Those are being used by Webkit to annotate
+// atomic operations, which should be handled by ThreadSanitizer correctly.
 void INTERFACE_ATTRIBUTE WTFAnnotateHappensBefore(char *f, int l, uptr addr) {
   SCOPED_ANNOTATION(AnnotateHappensBefore);
 }
@@ -368,6 +430,7 @@ void INTERFACE_ATTRIBUTE WTFAnnotateHappensAfter(char *f, int l, uptr addr) {
 void INTERFACE_ATTRIBUTE WTFAnnotateBenignRaceSized(
     char *f, int l, uptr mem, uptr sz, char *desc) {
   SCOPED_ANNOTATION(AnnotateBenignRaceSized);
+  BenignRaceImpl(f, l, mem, 1, desc);
 }
 
 int INTERFACE_ATTRIBUTE RunningOnValgrind() {
diff --git a/lib/tsan/rtl/tsan_interface_ann.h b/lib/tsan/rtl/tsan_interface_ann.h
index ed809073327e4..8e45328e7ec16 100644
--- a/lib/tsan/rtl/tsan_interface_ann.h
+++ b/lib/tsan/rtl/tsan_interface_ann.h
@@ -14,7 +14,7 @@
 #ifndef TSAN_INTERFACE_ANN_H
 #define TSAN_INTERFACE_ANN_H
 
-#include <sanitizer/common_interface_defs.h>
+#include <sanitizer_common/sanitizer_internal_defs.h>
 
 // This header should NOT include any other headers.
 // All functions in this header are extern "C" and start with __tsan_.
diff --git a/lib/tsan/rtl/tsan_interface_atomic.cc b/lib/tsan/rtl/tsan_interface_atomic.cc
index a9d75e5bf76ca..80266969849ae 100644
--- a/lib/tsan/rtl/tsan_interface_atomic.cc
+++ b/lib/tsan/rtl/tsan_interface_atomic.cc
@@ -20,25 +20,42 @@
 // http://www.hpl.hp.com/personal/Hans_Boehm/c++mm/
 
 #include "sanitizer_common/sanitizer_placement_new.h"
+#include "sanitizer_common/sanitizer_stacktrace.h"
 #include "tsan_interface_atomic.h"
 #include "tsan_flags.h"
 #include "tsan_rtl.h"
 
 using namespace __tsan;  // NOLINT
 
+#define SCOPED_ATOMIC(func, ...) \
+    const uptr callpc = (uptr)__builtin_return_address(0); \
+    uptr pc = __sanitizer::StackTrace::GetCurrentPc(); \
+    pc = __sanitizer::StackTrace::GetPreviousInstructionPc(pc); \
+    mo = ConvertOrder(mo); \
+    mo = flags()->force_seq_cst_atomics ? (morder)mo_seq_cst : mo; \
+    ThreadState *const thr = cur_thread(); \
+    AtomicStatInc(thr, sizeof(*a), mo, StatAtomic##func); \
+    ScopedAtomic sa(thr, callpc, __FUNCTION__); \
+    return Atomic##func(thr, pc, __VA_ARGS__); \
+/**/
+
 class ScopedAtomic {
  public:
   ScopedAtomic(ThreadState *thr, uptr pc, const char *func)
       : thr_(thr) {
-    CHECK_EQ(thr_->in_rtl, 1);  // 1 due to our own ScopedInRtl member.
+    CHECK_EQ(thr_->in_rtl, 0);
+    ProcessPendingSignals(thr);
+    FuncEntry(thr_, pc);
     DPrintf("#%d: %s\n", thr_->tid, func);
+    thr_->in_rtl++;
   }
   ~ScopedAtomic() {
-    CHECK_EQ(thr_->in_rtl, 1);
+    thr_->in_rtl--;
+    CHECK_EQ(thr_->in_rtl, 0);
+    FuncExit(thr_);
   }
  private:
   ThreadState *thr_;
-  ScopedInRtl in_rtl_;
 };
 
 // Some shortcuts.
@@ -212,16 +229,19 @@ a128 func_cas(volatile a128 *v, a128 cmp, a128 xch) {
 }
 #endif
 
-#define SCOPED_ATOMIC(func, ...) \
-    mo = ConvertOrder(mo); \
-    mo = flags()->force_seq_cst_atomics ? (morder)mo_seq_cst : mo; \
-    ThreadState *const thr = cur_thread(); \
-    ProcessPendingSignals(thr); \
-    const uptr pc = (uptr)__builtin_return_address(0); \
-    AtomicStatInc(thr, sizeof(*a), mo, StatAtomic##func); \
-    ScopedAtomic sa(thr, pc, __FUNCTION__); \
-    return Atomic##func(thr, pc, __VA_ARGS__); \
-/**/
+template<typename T>
+static int SizeLog() {
+  if (sizeof(T) <= 1)
+    return kSizeLog1;
+  else if (sizeof(T) <= 2)
+    return kSizeLog2;
+  else if (sizeof(T) <= 4)
+    return kSizeLog4;
+  else
+    return kSizeLog8;
+  // For 16-byte atomics we also use 8-byte memory access,
+  // this leads to false negatives only in very obscure cases.
+}
 
 template<typename T>
 static T AtomicLoad(ThreadState *thr, uptr pc, const volatile T *a,
@@ -229,14 +249,17 @@ static T AtomicLoad(ThreadState *thr, uptr pc, const volatile T *a,
   CHECK(IsLoadOrder(mo));
   // This fast-path is critical for performance.
   // Assume the access is atomic.
-  if (!IsAcquireOrder(mo) && sizeof(T) <= sizeof(a))
+  if (!IsAcquireOrder(mo) && sizeof(T) <= sizeof(a)) {
+    MemoryReadAtomic(thr, pc, (uptr)a, SizeLog<T>());
     return *a;
+  }
   SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, (uptr)a, false);
   thr->clock.set(thr->tid, thr->fast_state.epoch());
   thr->clock.acquire(&s->clock);
   T v = *a;
   s->mtx.ReadUnlock();
   __sync_synchronize();
+  MemoryReadAtomic(thr, pc, (uptr)a, SizeLog<T>());
   return v;
 }
 
@@ -244,6 +267,7 @@ template<typename T>
 static void AtomicStore(ThreadState *thr, uptr pc, volatile T *a, T v,
     morder mo) {
   CHECK(IsStoreOrder(mo));
+  MemoryWriteAtomic(thr, pc, (uptr)a, SizeLog<T>());
   // This fast-path is critical for performance.
   // Assume the access is atomic.
   // Strictly saying even relaxed store cuts off release sequence,
@@ -265,16 +289,21 @@ static void AtomicStore(ThreadState *thr, uptr pc, volatile T *a, T v,
 
 template<typename T, T (*F)(volatile T *v, T op)>
 static T AtomicRMW(ThreadState *thr, uptr pc, volatile T *a, T v, morder mo) {
-  SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, (uptr)a, true);
-  thr->clock.set(thr->tid, thr->fast_state.epoch());
-  if (IsAcqRelOrder(mo))
-    thr->clock.acq_rel(&s->clock);
-  else if (IsReleaseOrder(mo))
-    thr->clock.release(&s->clock);
-  else if (IsAcquireOrder(mo))
-    thr->clock.acquire(&s->clock);
+  MemoryWriteAtomic(thr, pc, (uptr)a, SizeLog<T>());
+  SyncVar *s = 0;
+  if (mo != mo_relaxed) {
+    s = CTX()->synctab.GetOrCreateAndLock(thr, pc, (uptr)a, true);
+    thr->clock.set(thr->tid, thr->fast_state.epoch());
+    if (IsAcqRelOrder(mo))
+      thr->clock.acq_rel(&s->clock);
+    else if (IsReleaseOrder(mo))
+      thr->clock.release(&s->clock);
+    else if (IsAcquireOrder(mo))
+      thr->clock.acquire(&s->clock);
+  }
   v = F(a, v);
-  s->mtx.Unlock();
+  if (s)
+    s->mtx.Unlock();
   return v;
 }
 
@@ -324,17 +353,22 @@ template<typename T>
 static bool AtomicCAS(ThreadState *thr, uptr pc,
     volatile T *a, T *c, T v, morder mo, morder fmo) {
   (void)fmo;  // Unused because llvm does not pass it yet.
-  SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, (uptr)a, true);
-  thr->clock.set(thr->tid, thr->fast_state.epoch());
-  if (IsAcqRelOrder(mo))
-    thr->clock.acq_rel(&s->clock);
-  else if (IsReleaseOrder(mo))
-    thr->clock.release(&s->clock);
-  else if (IsAcquireOrder(mo))
-    thr->clock.acquire(&s->clock);
+  MemoryWriteAtomic(thr, pc, (uptr)a, SizeLog<T>());
+  SyncVar *s = 0;
+  if (mo != mo_relaxed) {
+    s = CTX()->synctab.GetOrCreateAndLock(thr, pc, (uptr)a, true);
+    thr->clock.set(thr->tid, thr->fast_state.epoch());
+    if (IsAcqRelOrder(mo))
+      thr->clock.acq_rel(&s->clock);
+    else if (IsReleaseOrder(mo))
+      thr->clock.release(&s->clock);
+    else if (IsAcquireOrder(mo))
+      thr->clock.acquire(&s->clock);
+  }
   T cc = *c;
   T pr = func_cas(a, cc, v);
-  s->mtx.Unlock();
+  if (s)
+    s->mtx.Unlock();
   if (pr == cc)
     return true;
   *c = pr;
diff --git a/lib/tsan/rtl/tsan_interface_inl.h b/lib/tsan/rtl/tsan_interface_inl.h
index 8a92155d57ef8..0187e49d96e57 100644
--- a/lib/tsan/rtl/tsan_interface_inl.h
+++ b/lib/tsan/rtl/tsan_interface_inl.h
@@ -19,41 +19,53 @@
 using namespace __tsan;  // NOLINT
 
 void __tsan_read1(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 0, 0);
+  MemoryRead(cur_thread(), CALLERPC, (uptr)addr, kSizeLog1);
 }
 
 void __tsan_read2(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 1, 0);
+  MemoryRead(cur_thread(), CALLERPC, (uptr)addr, kSizeLog2);
 }
 
 void __tsan_read4(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 2, 0);
+  MemoryRead(cur_thread(), CALLERPC, (uptr)addr, kSizeLog4);
 }
 
 void __tsan_read8(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 3, 0);
+  MemoryRead(cur_thread(), CALLERPC, (uptr)addr, kSizeLog8);
 }
 
 void __tsan_write1(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 0, 1);
+  MemoryWrite(cur_thread(), CALLERPC, (uptr)addr, kSizeLog1);
 }
 
 void __tsan_write2(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 1, 1);
+  MemoryWrite(cur_thread(), CALLERPC, (uptr)addr, kSizeLog2);
 }
 
 void __tsan_write4(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 2, 1);
+  MemoryWrite(cur_thread(), CALLERPC, (uptr)addr, kSizeLog4);
 }
 
 void __tsan_write8(void *addr) {
-  MemoryAccess(cur_thread(), CALLERPC, (uptr)addr, 3, 1);
+  MemoryWrite(cur_thread(), CALLERPC, (uptr)addr, kSizeLog8);
 }
 
 void __tsan_vptr_update(void **vptr_p, void *new_val) {
   CHECK_EQ(sizeof(vptr_p), 8);
-  if (*vptr_p != new_val)
-    MemoryAccess(cur_thread(), CALLERPC, (uptr)vptr_p, 3, 1);
+  if (*vptr_p != new_val) {
+    ThreadState *thr = cur_thread();
+    thr->is_vptr_access = true;
+    MemoryWrite(thr, CALLERPC, (uptr)vptr_p, kSizeLog8);
+    thr->is_vptr_access = false;
+  }
+}
+
+void __tsan_vptr_read(void **vptr_p) {
+  CHECK_EQ(sizeof(vptr_p), 8);
+  ThreadState *thr = cur_thread();
+  thr->is_vptr_access = true;
+  MemoryRead(thr, CALLERPC, (uptr)vptr_p, kSizeLog8);
+  thr->is_vptr_access = false;
 }
 
 void __tsan_func_entry(void *pc) {
diff --git a/lib/tsan/rtl/tsan_interface_java.cc b/lib/tsan/rtl/tsan_interface_java.cc
index e425c75800be8..71e0747c3646f 100644
--- a/lib/tsan/rtl/tsan_interface_java.cc
+++ b/lib/tsan/rtl/tsan_interface_java.cc
@@ -152,7 +152,7 @@ SyncVar* GetAndRemoveJavaSync(ThreadState *thr, uptr pc, uptr addr) {
   return 0;
 }
 
-}  // namespace __tsan {
+}  // namespace __tsan
 
 #define SCOPED_JAVA_FUNC(func) \
   ThreadState *thr = cur_thread(); \
@@ -271,6 +271,7 @@ void __tsan_java_mutex_lock(jptr addr) {
   CHECK_GE(addr, jctx->heap_begin);
   CHECK_LT(addr, jctx->heap_begin + jctx->heap_size);
 
+  MutexCreate(thr, pc, addr, true, true, true);
   MutexLock(thr, pc, addr);
 }
 
@@ -291,6 +292,7 @@ void __tsan_java_mutex_read_lock(jptr addr) {
   CHECK_GE(addr, jctx->heap_begin);
   CHECK_LT(addr, jctx->heap_begin + jctx->heap_size);
 
+  MutexCreate(thr, pc, addr, true, true, true);
   MutexReadLock(thr, pc, addr);
 }
 
@@ -303,3 +305,25 @@ void __tsan_java_mutex_read_unlock(jptr addr) {
 
   MutexReadUnlock(thr, pc, addr);
 }
+
+void __tsan_java_mutex_lock_rec(jptr addr, int rec) {
+  SCOPED_JAVA_FUNC(__tsan_java_mutex_lock_rec);
+  DPrintf("#%d: java_mutex_lock_rec(%p, %d)\n", thr->tid, addr, rec);
+  CHECK_NE(jctx, 0);
+  CHECK_GE(addr, jctx->heap_begin);
+  CHECK_LT(addr, jctx->heap_begin + jctx->heap_size);
+  CHECK_GT(rec, 0);
+
+  MutexCreate(thr, pc, addr, true, true, true);
+  MutexLock(thr, pc, addr, rec);
+}
+
+int __tsan_java_mutex_unlock_rec(jptr addr) {
+  SCOPED_JAVA_FUNC(__tsan_java_mutex_unlock_rec);
+  DPrintf("#%d: java_mutex_unlock_rec(%p)\n", thr->tid, addr);
+  CHECK_NE(jctx, 0);
+  CHECK_GE(addr, jctx->heap_begin);
+  CHECK_LT(addr, jctx->heap_begin + jctx->heap_size);
+
+  return MutexUnlock(thr, pc, addr, true);
+}
diff --git a/lib/tsan/rtl/tsan_interface_java.h b/lib/tsan/rtl/tsan_interface_java.h
index 241483aaa015b..9ac78e074bbe0 100644
--- a/lib/tsan/rtl/tsan_interface_java.h
+++ b/lib/tsan/rtl/tsan_interface_java.h
@@ -55,8 +55,7 @@ void __tsan_java_move(jptr src, jptr dst, jptr size) INTERFACE_ATTRIBUTE;
 
 // Mutex lock.
 // Addr is any unique address associated with the mutex.
-// Must not be called on recursive reentry.
-// Object.wait() is handled as a pair of unlock/lock.
+// Can be called on recursive reentry.
 void __tsan_java_mutex_lock(jptr addr) INTERFACE_ATTRIBUTE;
 // Mutex unlock.
 void __tsan_java_mutex_unlock(jptr addr) INTERFACE_ATTRIBUTE;
@@ -64,6 +63,16 @@ void __tsan_java_mutex_unlock(jptr addr) INTERFACE_ATTRIBUTE;
 void __tsan_java_mutex_read_lock(jptr addr) INTERFACE_ATTRIBUTE;
 // Mutex read unlock.
 void __tsan_java_mutex_read_unlock(jptr addr) INTERFACE_ATTRIBUTE;
+// Recursive mutex lock, intended for handling of Object.wait().
+// The 'rec' value must be obtained from the previous
+// __tsan_java_mutex_unlock_rec().
+void __tsan_java_mutex_lock_rec(jptr addr, int rec) INTERFACE_ATTRIBUTE;
+// Recursive mutex unlock, intended for handling of Object.wait().
+// The return value says how many times this thread called lock()
+// w/o a pairing unlock() (i.e. how many recursive levels it unlocked).
+// It must be passed back to __tsan_java_mutex_lock_rec() to restore
+// the same recursion level.
+int __tsan_java_mutex_unlock_rec(jptr addr) INTERFACE_ATTRIBUTE;
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/lib/tsan/rtl/tsan_md5.cc b/lib/tsan/rtl/tsan_md5.cc
index c9d671f5b5993..66e8240431537 100644
--- a/lib/tsan/rtl/tsan_md5.cc
+++ b/lib/tsan/rtl/tsan_md5.cc
@@ -242,4 +242,4 @@ MD5Hash md5_hash(const void *data, uptr size) {
   MD5_Final((unsigned char*)&res.hash[0], &ctx);
   return res;
 }
-}
+}  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_mman.cc b/lib/tsan/rtl/tsan_mman.cc
index 82f7105d60db8..b6671b1abf09e 100644
--- a/lib/tsan/rtl/tsan_mman.cc
+++ b/lib/tsan/rtl/tsan_mman.cc
@@ -29,6 +29,41 @@ extern "C" void WEAK __tsan_free_hook(void *ptr) {
 
 namespace __tsan {
 
+COMPILER_CHECK(sizeof(MBlock) == 16);
+
+void MBlock::Lock() {
+  atomic_uintptr_t *a = reinterpret_cast<atomic_uintptr_t*>(this);
+  uptr v = atomic_load(a, memory_order_relaxed);
+  for (int iter = 0;; iter++) {
+    if (v & 1) {
+      if (iter < 10)
+        proc_yield(20);
+      else
+        internal_sched_yield();
+      v = atomic_load(a, memory_order_relaxed);
+      continue;
+    }
+    if (atomic_compare_exchange_weak(a, &v, v | 1, memory_order_acquire))
+      break;
+  }
+}
+
+void MBlock::Unlock() {
+  atomic_uintptr_t *a = reinterpret_cast<atomic_uintptr_t*>(this);
+  uptr v = atomic_load(a, memory_order_relaxed);
+  DCHECK(v & 1);
+  atomic_store(a, v & ~1, memory_order_relaxed);
+}
+
+struct MapUnmapCallback {
+  void OnMap(uptr p, uptr size) const { }
+  void OnUnmap(uptr p, uptr size) const {
+    // We are about to unmap a chunk of user memory.
+    // Mark the corresponding shadow memory as not needed.
+    DontNeedShadowFor(p, size);
+  }
+};
+
 static char allocator_placeholder[sizeof(Allocator)] ALIGNED(64);
 Allocator *allocator() {
   return reinterpret_cast<Allocator*>(&allocator_placeholder);
@@ -38,8 +73,16 @@ void InitializeAllocator() {
   allocator()->Init();
 }
 
-void AlloctorThreadFinish(ThreadState *thr) {
-  allocator()->SwallowCache(&thr->alloc_cache);
+void AllocatorThreadStart(ThreadState *thr) {
+  allocator()->InitCache(&thr->alloc_cache);
+}
+
+void AllocatorThreadFinish(ThreadState *thr) {
+  allocator()->DestroyCache(&thr->alloc_cache);
+}
+
+void AllocatorPrintStats() {
+  allocator()->PrintStats();
 }
 
 static void SignalUnsafeCall(ThreadState *thr, uptr pc) {
@@ -48,7 +91,7 @@ static void SignalUnsafeCall(ThreadState *thr, uptr pc) {
   Context *ctx = CTX();
   StackTrace stack;
   stack.ObtainCurrent(thr, pc);
-  Lock l(&ctx->thread_mtx);
+  ThreadRegistryLock l(ctx->thread_registry);
   ScopedReport rep(ReportTypeSignalUnsafe);
   if (!IsFiredSuppression(ctx, rep, stack)) {
     rep.AddStack(&stack);
@@ -58,17 +101,15 @@ static void SignalUnsafeCall(ThreadState *thr, uptr pc) {
 
 void *user_alloc(ThreadState *thr, uptr pc, uptr sz, uptr align) {
   CHECK_GT(thr->in_rtl, 0);
+  if ((sz >= (1ull << 40)) || (align >= (1ull << 40)))
+    return 0;
   void *p = allocator()->Allocate(&thr->alloc_cache, sz, align);
   if (p == 0)
     return 0;
   MBlock *b = new(allocator()->GetMetaData(p)) MBlock;
-  b->size = sz;
-  b->head = 0;
-  b->alloc_tid = thr->unique_id;
-  b->alloc_stack_id = CurrentStackId(thr, pc);
-  if (CTX() && CTX()->initialized) {
+  b->Init(sz, thr->tid, CurrentStackId(thr, pc));
+  if (CTX() && CTX()->initialized)
     MemoryRangeImitateWrite(thr, pc, (uptr)p, sz);
-  }
   DPrintf("#%d: alloc(%zu) = %p\n", thr->tid, sz, p);
   SignalUnsafeCall(thr, pc);
   return p;
@@ -79,9 +120,9 @@ void user_free(ThreadState *thr, uptr pc, void *p) {
   CHECK_NE(p, (void*)0);
   DPrintf("#%d: free(%p)\n", thr->tid, p);
   MBlock *b = (MBlock*)allocator()->GetMetaData(p);
-  if (b->head)   {
-    Lock l(&b->mtx);
-    for (SyncVar *s = b->head; s;) {
+  if (b->ListHead()) {
+    MBlock::ScopedLock l(b);
+    for (SyncVar *s = b->ListHead(); s;) {
       SyncVar *res = s;
       s = s->next;
       StatInc(thr, StatSyncDestroyed);
@@ -89,12 +130,10 @@ void user_free(ThreadState *thr, uptr pc, void *p) {
       res->mtx.Unlock();
       DestroyAndFree(res);
     }
-    b->head = 0;
+    b->ListReset();
   }
-  if (CTX() && CTX()->initialized && thr->in_rtl == 1) {
-    MemoryRangeFreed(thr, pc, (uptr)p, b->size);
-  }
-  b->~MBlock();
+  if (CTX() && CTX()->initialized && thr->in_rtl == 1)
+    MemoryRangeFreed(thr, pc, (uptr)p, b->Size());
   allocator()->Deallocate(&thr->alloc_cache, p);
   SignalUnsafeCall(thr, pc);
 }
@@ -110,20 +149,29 @@ void *user_realloc(ThreadState *thr, uptr pc, void *p, uptr sz) {
       return 0;
     if (p) {
       MBlock *b = user_mblock(thr, p);
-      internal_memcpy(p2, p, min(b->size, sz));
+      CHECK_NE(b, 0);
+      internal_memcpy(p2, p, min(b->Size(), sz));
     }
   }
-  if (p) {
+  if (p)
     user_free(thr, pc, p);
-  }
   return p2;
 }
 
+uptr user_alloc_usable_size(ThreadState *thr, uptr pc, void *p) {
+  CHECK_GT(thr->in_rtl, 0);
+  if (p == 0)
+    return 0;
+  MBlock *b = (MBlock*)allocator()->GetMetaData(p);
+  return b ? b->Size() : 0;
+}
+
 MBlock *user_mblock(ThreadState *thr, void *p) {
-  CHECK_NE(p, (void*)0);
+  CHECK_NE(p, 0);
   Allocator *a = allocator();
   void *b = a->GetBlockBegin(p);
-  CHECK_NE(b, 0);
+  if (b == 0)
+    return 0;
   return (MBlock*)a->GetMetaData(b);
 }
 
@@ -164,3 +212,54 @@ void internal_free(void *p) {
 }
 
 }  // namespace __tsan
+
+using namespace __tsan;
+
+extern "C" {
+uptr __tsan_get_current_allocated_bytes() {
+  u64 stats[AllocatorStatCount];
+  allocator()->GetStats(stats);
+  u64 m = stats[AllocatorStatMalloced];
+  u64 f = stats[AllocatorStatFreed];
+  return m >= f ? m - f : 1;
+}
+
+uptr __tsan_get_heap_size() {
+  u64 stats[AllocatorStatCount];
+  allocator()->GetStats(stats);
+  u64 m = stats[AllocatorStatMmapped];
+  u64 f = stats[AllocatorStatUnmapped];
+  return m >= f ? m - f : 1;
+}
+
+uptr __tsan_get_free_bytes() {
+  return 1;
+}
+
+uptr __tsan_get_unmapped_bytes() {
+  return 1;
+}
+
+uptr __tsan_get_estimated_allocated_size(uptr size) {
+  return size;
+}
+
+bool __tsan_get_ownership(void *p) {
+  return allocator()->GetBlockBegin(p) != 0;
+}
+
+uptr __tsan_get_allocated_size(void *p) {
+  if (p == 0)
+    return 0;
+  p = allocator()->GetBlockBegin(p);
+  if (p == 0)
+    return 0;
+  MBlock *b = (MBlock*)allocator()->GetMetaData(p);
+  return b->Size();
+}
+
+void __tsan_on_thread_idle() {
+  ThreadState *thr = cur_thread();
+  allocator()->SwallowCache(&thr->alloc_cache);
+}
+}  // extern "C"
diff --git a/lib/tsan/rtl/tsan_mman.h b/lib/tsan/rtl/tsan_mman.h
index 5cf00eac8d030..19d555437f3ec 100644
--- a/lib/tsan/rtl/tsan_mman.h
+++ b/lib/tsan/rtl/tsan_mman.h
@@ -20,7 +20,9 @@ namespace __tsan {
 const uptr kDefaultAlignment = 16;
 
 void InitializeAllocator();
-void AlloctorThreadFinish(ThreadState *thr);
+void AllocatorThreadStart(ThreadState *thr);
+void AllocatorThreadFinish(ThreadState *thr);
+void AllocatorPrintStats();
 
 // For user allocations.
 void *user_alloc(ThreadState *thr, uptr pc, uptr sz,
@@ -29,6 +31,7 @@ void *user_alloc(ThreadState *thr, uptr pc, uptr sz,
 void user_free(ThreadState *thr, uptr pc, void *p);
 void *user_realloc(ThreadState *thr, uptr pc, void *p, uptr sz);
 void *user_alloc_aligned(ThreadState *thr, uptr pc, uptr sz, uptr align);
+uptr user_alloc_usable_size(ThreadState *thr, uptr pc, void *p);
 // Given the pointer p into a valid allocated block,
 // returns the descriptor of the block.
 MBlock *user_mblock(ThreadState *thr, void *p);
@@ -60,6 +63,7 @@ enum MBlockType {
   MBlockExpectRace,
   MBlockSignal,
   MBlockFD,
+  MBlockJmpBuf,
 
   // This must be the last.
   MBlockTypeCount
diff --git a/lib/tsan/rtl/tsan_mutex.cc b/lib/tsan/rtl/tsan_mutex.cc
index 335ca2211d132..a92fd90fd9c15 100644
--- a/lib/tsan/rtl/tsan_mutex.cc
+++ b/lib/tsan/rtl/tsan_mutex.cc
@@ -31,8 +31,8 @@ static MutexType CanLockTab[MutexTypeCount][MutexTypeCount] = {
   /*0  MutexTypeInvalid*/     {},
   /*1  MutexTypeTrace*/       {MutexTypeLeaf},
   /*2  MutexTypeThreads*/     {MutexTypeReport},
-  /*3  MutexTypeReport*/      {MutexTypeSyncTab, MutexTypeMBlock,
-                               MutexTypeJavaMBlock},
+  /*3  MutexTypeReport*/      {MutexTypeSyncTab, MutexTypeSyncVar,
+                               MutexTypeMBlock, MutexTypeJavaMBlock},
   /*4  MutexTypeSyncVar*/     {},
   /*5  MutexTypeSyncTab*/     {MutexTypeSyncVar},
   /*6  MutexTypeSlab*/        {MutexTypeLeaf},
diff --git a/lib/tsan/rtl/tsan_mutexset.h b/lib/tsan/rtl/tsan_mutexset.h
index 09223ff6cc481..eebfd4d70a142 100644
--- a/lib/tsan/rtl/tsan_mutexset.h
+++ b/lib/tsan/rtl/tsan_mutexset.h
@@ -22,7 +22,7 @@ class MutexSet {
  public:
   // Holds limited number of mutexes.
   // The oldest mutexes are discarded on overflow.
-  static const uptr kMaxSize = 64;
+  static const uptr kMaxSize = 16;
   struct Desc {
     u64 id;
     u64 epoch;
diff --git a/lib/tsan/rtl/tsan_platform.h b/lib/tsan/rtl/tsan_platform.h
index c859c3e85b198..666b4d0c482f2 100644
--- a/lib/tsan/rtl/tsan_platform.h
+++ b/lib/tsan/rtl/tsan_platform.h
@@ -37,9 +37,9 @@ C++ COMPAT linux memory layout:
 Go linux and darwin memory layout:
 0000 0000 0000 - 0000 1000 0000: executable
 0000 1000 0000 - 00f8 0000 0000: -
-00f8 0000 0000 - 0118 0000 0000: heap
-0118 0000 0000 - 1000 0000 0000: -
-1000 0000 0000 - 1460 0000 0000: shadow
+00c0 0000 0000 - 00e0 0000 0000: heap
+00e0 0000 0000 - 1000 0000 0000: -
+1000 0000 0000 - 1380 0000 0000: shadow
 1460 0000 0000 - 6000 0000 0000: -
 6000 0000 0000 - 6200 0000 0000: traces
 6200 0000 0000 - 7fff ffff ffff: -
@@ -47,8 +47,8 @@ Go linux and darwin memory layout:
 Go windows memory layout:
 0000 0000 0000 - 0000 1000 0000: executable
 0000 1000 0000 - 00f8 0000 0000: -
-00f8 0000 0000 - 0118 0000 0000: heap
-0118 0000 0000 - 0100 0000 0000: -
+00c0 0000 0000 - 00e0 0000 0000: heap
+00e0 0000 0000 - 0100 0000 0000: -
 0100 0000 0000 - 0560 0000 0000: shadow
 0560 0000 0000 - 0760 0000 0000: traces
 0760 0000 0000 - 07ff ffff ffff: -
@@ -65,11 +65,11 @@ namespace __tsan {
 
 #if defined(TSAN_GO)
 static const uptr kLinuxAppMemBeg = 0x000000000000ULL;
-static const uptr kLinuxAppMemEnd = 0x00fcffffffffULL;
-# if defined(_WIN32)
+static const uptr kLinuxAppMemEnd = 0x04dfffffffffULL;
+# if SANITIZER_WINDOWS
 static const uptr kLinuxShadowMsk = 0x010000000000ULL;
 # else
-static const uptr kLinuxShadowMsk = 0x100000000000ULL;
+static const uptr kLinuxShadowMsk = 0x200000000000ULL;
 # endif
 // TSAN_COMPAT_SHADOW is intended for COMPAT virtual memory layout,
 // when memory addresses are of the 0x2axxxxxxxxxx form.
@@ -84,7 +84,7 @@ static const uptr kLinuxAppMemEnd = 0x7fffffffffffULL;
 
 static const uptr kLinuxAppMemMsk = 0x7c0000000000ULL;
 
-#if defined(_WIN32)
+#if SANITIZER_WINDOWS
 const uptr kTraceMemBegin = 0x056000000000ULL;
 #else
 const uptr kTraceMemBegin = 0x600000000000ULL;
@@ -132,13 +132,19 @@ static inline uptr AlternativeAddress(uptr addr) {
 #endif
 }
 
-uptr GetShadowMemoryConsumption();
 void FlushShadowMemory();
+void WriteMemoryProfile(char *buf, uptr buf_size);
 
 const char *InitializePlatform();
 void FinalizePlatform();
-uptr ALWAYS_INLINE INLINE GetThreadTrace(int tid) {
-  uptr p = kTraceMemBegin + (uptr)tid * kTraceSize * sizeof(Event);
+uptr ALWAYS_INLINE GetThreadTrace(int tid) {
+  uptr p = kTraceMemBegin + (uptr)(tid * 2) * kTraceSize * sizeof(Event);
+  DCHECK_LT(p, kTraceMemBegin + kTraceMemSize);
+  return p;
+}
+
+uptr ALWAYS_INLINE GetThreadTraceHeader(int tid) {
+  uptr p = kTraceMemBegin + (uptr)(tid * 2 + 1) * kTraceSize * sizeof(Event);
   DCHECK_LT(p, kTraceMemBegin + kTraceMemSize);
   return p;
 }
@@ -148,9 +154,7 @@ void internal_start_thread(void(*func)(void*), void *arg);
 // Says whether the addr relates to a global var.
 // Guesses with high probability, may yield both false positives and negatives.
 bool IsGlobalVar(uptr addr);
-uptr GetTlsSize();
-void GetThreadStackAndTls(bool main, uptr *stk_addr, uptr *stk_size,
-                          uptr *tls_addr, uptr *tls_size);
+int ExtractResolvFDs(void *state, int *fds, int nfd);
 
 }  // namespace __tsan
 
diff --git a/lib/tsan/rtl/tsan_platform_linux.cc b/lib/tsan/rtl/tsan_platform_linux.cc
index 6cc4249751259..a0d71e8589d65 100644
--- a/lib/tsan/rtl/tsan_platform_linux.cc
+++ b/lib/tsan/rtl/tsan_platform_linux.cc
@@ -12,7 +12,9 @@
 // Linux-specific code.
 //===----------------------------------------------------------------------===//
 
-#ifdef __linux__
+
+#include "sanitizer_common/sanitizer_platform.h"
+#if SANITIZER_LINUX
 
 #include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_libc.h"
@@ -21,7 +23,6 @@
 #include "tsan_rtl.h"
 #include "tsan_flags.h"
 
-#include <asm/prctl.h>
 #include <fcntl.h>
 #include <pthread.h>
 #include <signal.h>
@@ -40,11 +41,16 @@
 #include <errno.h>
 #include <sched.h>
 #include <dlfcn.h>
+#define __need_res_state
+#include <resolv.h>
+#include <malloc.h>
 
-extern "C" int arch_prctl(int code, __sanitizer::uptr *addr);
+extern "C" struct mallinfo __libc_mallinfo();
 
 namespace __tsan {
 
+const uptr kPageSize = 4096;
+
 #ifndef TSAN_GO
 ScopedInRtl::ScopedInRtl()
     : thr_(cur_thread()) {
@@ -66,8 +72,75 @@ ScopedInRtl::~ScopedInRtl() {
 }
 #endif
 
-uptr GetShadowMemoryConsumption() {
-  return 0;
+static bool ishex(char c) {
+  return (c >= '0' && c <= '9')
+      || (c >= 'a' && c <= 'f');
+}
+
+static uptr readhex(const char *p) {
+  uptr v = 0;
+  for (; ishex(p[0]); p++) {
+    if (p[0] >= '0' && p[0] <= '9')
+      v = v * 16 + p[0] - '0';
+    else
+      v = v * 16 + p[0] - 'a' + 10;
+  }
+  return v;
+}
+
+static uptr readdec(const char *p) {
+  uptr v = 0;
+  for (; p[0] >= '0' && p[0] <= '9' ; p++)
+    v = v * 10 + p[0] - '0';
+  return v;
+}
+
+void WriteMemoryProfile(char *buf, uptr buf_size) {
+  char *smaps = 0;
+  uptr smaps_cap = 0;
+  uptr smaps_len = ReadFileToBuffer("/proc/self/smaps",
+      &smaps, &smaps_cap, 64<<20);
+  uptr mem[6] = {};
+  uptr total = 0;
+  uptr start = 0;
+  bool file = false;
+  const char *pos = smaps;
+  while (pos < smaps + smaps_len) {
+    if (ishex(pos[0])) {
+      start = readhex(pos);
+      for (; *pos != '/' && *pos > '\n'; pos++) {}
+      file = *pos == '/';
+    } else if (internal_strncmp(pos, "Rss:", 4) == 0) {
+      for (; *pos < '0' || *pos > '9'; pos++) {}
+      uptr rss = readdec(pos) * 1024;
+      total += rss;
+      start >>= 40;
+      if (start < 0x10)  // shadow
+        mem[0] += rss;
+      else if (start >= 0x20 && start < 0x30)  // compat modules
+        mem[file ? 1 : 2] += rss;
+      else if (start >= 0x7e)  // modules
+        mem[file ? 1 : 2] += rss;
+      else if (start >= 0x60 && start < 0x62)  // traces
+        mem[3] += rss;
+      else if (start >= 0x7d && start < 0x7e)  // heap
+        mem[4] += rss;
+      else  // other
+        mem[5] += rss;
+    }
+    while (*pos++ != '\n') {}
+  }
+  UnmapOrDie(smaps, smaps_cap);
+  char *buf_pos = buf;
+  char *buf_end = buf + buf_size;
+  buf_pos += internal_snprintf(buf_pos, buf_end - buf_pos,
+      "RSS %zd MB: shadow:%zd file:%zd mmap:%zd trace:%zd heap:%zd other:%zd\n",
+      total >> 20, mem[0] >> 20, mem[1] >> 20, mem[2] >> 20,
+      mem[3] >> 20, mem[4] >> 20, mem[5] >> 20);
+  struct mallinfo mi = __libc_mallinfo();
+  buf_pos += internal_snprintf(buf_pos, buf_end - buf_pos,
+      "mallinfo: arena=%d mmap=%d fordblks=%d keepcost=%d\n",
+      mi.arena >> 20, mi.hblkhd >> 20, mi.fordblks >> 20, mi.keepcost >> 20);
 }
 
 void FlushShadowMemory() {
@@ -89,6 +162,63 @@ static void ProtectRange(uptr beg, uptr end) {
 #endif
 
 #ifndef TSAN_GO
+// Mark shadow for .rodata sections with the special kShadowRodata marker.
+// Accesses to .rodata can't race, so this saves time, memory and trace space.
+static void MapRodata() {
+  // First create temp file.
+  const char *tmpdir = GetEnv("TMPDIR");
+  if (tmpdir == 0)
+    tmpdir = GetEnv("TEST_TMPDIR");
+#ifdef P_tmpdir
+  if (tmpdir == 0)
+    tmpdir = P_tmpdir;
+#endif
+  if (tmpdir == 0)
+    return;
+  char filename[256];
+  internal_snprintf(filename, sizeof(filename), "%s/tsan.rodata.%d",
+                    tmpdir, (int)internal_getpid());
+  uptr openrv = internal_open(filename, O_RDWR | O_CREAT | O_EXCL, 0600);
+  if (internal_iserror(openrv))
+    return;
+  fd_t fd = openrv;
+  // Fill the file with kShadowRodata.
+  const uptr kMarkerSize = 512 * 1024 / sizeof(u64);
+  InternalScopedBuffer<u64> marker(kMarkerSize);
+  for (u64 *p = marker.data(); p < marker.data() + kMarkerSize; p++)
+    *p = kShadowRodata;
+  internal_write(fd, marker.data(), marker.size());
+  // Map the file into memory.
+  uptr page = internal_mmap(0, kPageSize, PROT_READ | PROT_WRITE,
+                            MAP_PRIVATE | MAP_ANONYMOUS, fd, 0);
+  if (internal_iserror(page)) {
+    internal_close(fd);
+    internal_unlink(filename);
+    return;
+  }
+  // Map the file into shadow of .rodata sections.
+  MemoryMappingLayout proc_maps(/*cache_enabled*/true);
+  uptr start, end, offset, prot;
+  char name[128];
+  while (proc_maps.Next(&start, &end, &offset, name, ARRAY_SIZE(name), &prot)) {
+    if (name[0] != 0 && name[0] != '['
+        && (prot & MemoryMappingLayout::kProtectionRead)
+        && (prot & MemoryMappingLayout::kProtectionExecute)
+        && !(prot & MemoryMappingLayout::kProtectionWrite)
+        && IsAppMem(start)) {
+      // Assume it's .rodata
+      char *shadow_start = (char*)MemToShadow(start);
+      char *shadow_end = (char*)MemToShadow(end);
+      for (char *p = shadow_start; p < shadow_end; p += marker.size()) {
+        internal_mmap(p, Min<uptr>(marker.size(), shadow_end - p),
+                      PROT_READ, MAP_PRIVATE | MAP_FIXED, fd, 0);
+      }
+    }
+  }
+  internal_close(fd);
+  internal_unlink(filename);
+}
+
 void InitializeShadowMemory() {
   uptr shadow = (uptr)MmapFixedNoReserve(kLinuxShadowBeg,
     kLinuxShadowEnd - kLinuxShadowBeg);
@@ -115,6 +245,8 @@ void InitializeShadowMemory() {
       kLinuxAppMemBeg, kLinuxAppMemEnd,
       (kLinuxAppMemEnd - kLinuxAppMemBeg) >> 30);
   DPrintf("stack        %zx\n", (uptr)&shadow);
+
+  MapRodata();
 }
 #endif
 
@@ -124,10 +256,11 @@ static uptr g_data_end;
 #ifndef TSAN_GO
 static void CheckPIE() {
   // Ensure that the binary is indeed compiled with -pie.
-  MemoryMappingLayout proc_maps;
+  MemoryMappingLayout proc_maps(true);
   uptr start, end;
   if (proc_maps.Next(&start, &end,
-                     /*offset*/0, /*filename*/0, /*filename_size*/0)) {
+                     /*offset*/0, /*filename*/0, /*filename_size*/0,
+                     /*protection*/0)) {
     if ((u64)start < kLinuxAppMemBeg) {
       Printf("FATAL: ThreadSanitizer can not mmap the shadow memory ("
              "something is mapped at 0x%zx < 0x%zx)\n",
@@ -140,11 +273,12 @@ static void CheckPIE() {
 }
 
 static void InitDataSeg() {
-  MemoryMappingLayout proc_maps;
+  MemoryMappingLayout proc_maps(true);
   uptr start, end, offset;
   char name[128];
   bool prev_is_data = false;
-  while (proc_maps.Next(&start, &end, &offset, name, ARRAY_SIZE(name))) {
+  while (proc_maps.Next(&start, &end, &offset, name, ARRAY_SIZE(name),
+                        /*protection*/ 0)) {
     DPrintf("%p-%p %p %s\n", start, end, offset, name);
     bool is_data = offset != 0 && name[0] != 0;
     // BSS may get merged with [heap] in /proc/self/maps. This is not very
@@ -163,27 +297,6 @@ static void InitDataSeg() {
   CHECK_LT((uptr)&g_data_start, g_data_end);
 }
 
-static uptr g_tls_size;
-
-#ifdef __i386__
-# define INTERNAL_FUNCTION __attribute__((regparm(3), stdcall))
-#else
-# define INTERNAL_FUNCTION
-#endif
-
-static int InitTlsSize() {
-  typedef void (*get_tls_func)(size_t*, size_t*) INTERNAL_FUNCTION;
-  get_tls_func get_tls;
-  void *get_tls_static_info_ptr = dlsym(RTLD_NEXT, "_dl_get_tls_static_info");
-  CHECK_EQ(sizeof(get_tls), sizeof(get_tls_static_info_ptr));
-  internal_memcpy(&get_tls, &get_tls_static_info_ptr,
-                  sizeof(get_tls_static_info_ptr));
-  CHECK_NE(get_tls, 0);
-  size_t tls_size = 0;
-  size_t tls_align = 0;
-  get_tls(&tls_size, &tls_align);
-  return tls_size;
-}
 #endif  // #ifndef TSAN_GO
 
 static rlim_t getlim(int res) {
@@ -238,57 +351,29 @@ const char *InitializePlatform() {
 
 #ifndef TSAN_GO
   CheckPIE();
-  g_tls_size = (uptr)InitTlsSize();
+  InitTlsSize();
   InitDataSeg();
 #endif
-  return getenv(kTsanOptionsEnv);
-}
-
-void FinalizePlatform() {
-  fflush(0);
+  return GetEnv(kTsanOptionsEnv);
 }
 
-uptr GetTlsSize() {
-#ifndef TSAN_GO
-  return g_tls_size;
-#else
-  return 0;
-#endif
+bool IsGlobalVar(uptr addr) {
+  return g_data_start && addr >= g_data_start && addr < g_data_end;
 }
 
-void GetThreadStackAndTls(bool main, uptr *stk_addr, uptr *stk_size,
-                          uptr *tls_addr, uptr *tls_size) {
 #ifndef TSAN_GO
-  arch_prctl(ARCH_GET_FS, tls_addr);
-  *tls_addr -= g_tls_size;
-  *tls_size = g_tls_size;
-
-  uptr stack_top, stack_bottom;
-  GetThreadStackTopAndBottom(main, &stack_top, &stack_bottom);
-  *stk_addr = stack_bottom;
-  *stk_size = stack_top - stack_bottom;
-
-  if (!main) {
-    // If stack and tls intersect, make them non-intersecting.
-    if (*tls_addr > *stk_addr && *tls_addr < *stk_addr + *stk_size) {
-      CHECK_GT(*tls_addr + *tls_size, *stk_addr);
-      CHECK_LE(*tls_addr + *tls_size, *stk_addr + *stk_size);
-      *stk_size -= *tls_size;
-      *tls_addr = *stk_addr + *stk_size;
-    }
+int ExtractResolvFDs(void *state, int *fds, int nfd) {
+  int cnt = 0;
+  __res_state *statp = (__res_state*)state;
+  for (int i = 0; i < MAXNS && cnt < nfd; i++) {
+    if (statp->_u._ext.nsaddrs[i] && statp->_u._ext.nssocks[i] != -1)
+      fds[cnt++] = statp->_u._ext.nssocks[i];
   }
-#else
-  *stk_addr = 0;
-  *stk_size = 0;
-  *tls_addr = 0;
-  *tls_size = 0;
-#endif
+  return cnt;
 }
+#endif
 
-bool IsGlobalVar(uptr addr) {
-  return g_data_start && addr >= g_data_start && addr < g_data_end;
-}
 
 }  // namespace __tsan
 
-#endif  // #ifdef __linux__
+#endif  // SANITIZER_LINUX
diff --git a/lib/tsan/rtl/tsan_platform_mac.cc b/lib/tsan/rtl/tsan_platform_mac.cc
index 183061d146382..99d4533a4fa26 100644
--- a/lib/tsan/rtl/tsan_platform_mac.cc
+++ b/lib/tsan/rtl/tsan_platform_mac.cc
@@ -12,7 +12,8 @@
 // Mac-specific code.
 //===----------------------------------------------------------------------===//
 
-#ifdef __APPLE__
+#include "sanitizer_common/sanitizer_platform.h"
+#if SANITIZER_MAC
 
 #include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_libc.h"
@@ -82,25 +83,13 @@ const char *InitializePlatform() {
     setrlimit(RLIMIT_CORE, (rlimit*)&lim);
   }
 
-  return getenv(kTsanOptionsEnv);
+  return GetEnv(kTsanOptionsEnv);
 }
 
 void FinalizePlatform() {
   fflush(0);
 }
 
-uptr GetTlsSize() {
-  return 0;
-}
-
-void GetThreadStackAndTls(bool main, uptr *stk_addr, uptr *stk_size,
-                          uptr *tls_addr, uptr *tls_size) {
-  *stk_addr = 0;
-  *stk_size = 0;
-  *tls_addr = 0;
-  *tls_size = 0;
-}
-
 }  // namespace __tsan
 
-#endif  // #ifdef __APPLE__
+#endif  // SANITIZER_MAC
diff --git a/lib/tsan/rtl/tsan_platform_windows.cc b/lib/tsan/rtl/tsan_platform_windows.cc
index f23e84e7875d5..711db72ce6847 100644
--- a/lib/tsan/rtl/tsan_platform_windows.cc
+++ b/lib/tsan/rtl/tsan_platform_windows.cc
@@ -12,7 +12,8 @@
 // Windows-specific code.
 //===----------------------------------------------------------------------===//
 
-#ifdef _WIN32
+#include "sanitizer_common/sanitizer_platform.h"
+#if SANITIZER_WINDOWS
 
 #include "tsan_platform.h"
 
@@ -34,25 +35,13 @@ void FlushShadowMemory() {
 }
 
 const char *InitializePlatform() {
-  return getenv(kTsanOptionsEnv);
+  return GetEnv(kTsanOptionsEnv);
 }
 
 void FinalizePlatform() {
   fflush(0);
 }
 
-uptr GetTlsSize() {
-  return 0;
-}
-
-void GetThreadStackAndTls(bool main, uptr *stk_addr, uptr *stk_size,
-                          uptr *tls_addr, uptr *tls_size) {
-  *stk_addr = 0;
-  *stk_size = 0;
-  *tls_addr = 0;
-  *tls_size = 0;
-}
-
 }  // namespace __tsan
 
-#endif  // #ifdef _WIN32
+#endif  // SANITIZER_WINDOWS
diff --git a/lib/tsan/rtl/tsan_report.cc b/lib/tsan/rtl/tsan_report.cc
index 056dc97387b90..c95c5c86be69c 100644
--- a/lib/tsan/rtl/tsan_report.cc
+++ b/lib/tsan/rtl/tsan_report.cc
@@ -22,7 +22,8 @@ ReportDesc::ReportDesc()
     , locs(MBlockReportLoc)
     , mutexes(MBlockReportMutex)
     , threads(MBlockReportThread)
-    , sleep() {
+    , sleep()
+    , count() {
 }
 
 ReportMop::ReportMop()
@@ -43,23 +44,22 @@ const char *thread_name(char *buf, int tid) {
   return buf;
 }
 
-static void PrintHeader(ReportType typ) {
-  Printf("WARNING: ThreadSanitizer: ");
-
+static const char *ReportTypeString(ReportType typ) {
   if (typ == ReportTypeRace)
-    Printf("data race");
-  else if (typ == ReportTypeUseAfterFree)
-    Printf("heap-use-after-free");
-  else if (typ == ReportTypeThreadLeak)
-    Printf("thread leak");
-  else if (typ == ReportTypeMutexDestroyLocked)
-    Printf("destroy of a locked mutex");
-  else if (typ == ReportTypeSignalUnsafe)
-    Printf("signal-unsafe call inside of a signal");
-  else if (typ == ReportTypeErrnoInSignal)
-    Printf("signal handler spoils errno");
-
-  Printf(" (pid=%d)\n", GetPid());
+    return "data race";
+  if (typ == ReportTypeVptrRace)
+    return "data race on vptr (ctor/dtor vs virtual call)";
+  if (typ == ReportTypeUseAfterFree)
+    return "heap-use-after-free";
+  if (typ == ReportTypeThreadLeak)
+    return "thread leak";
+  if (typ == ReportTypeMutexDestroyLocked)
+    return "destroy of a locked mutex";
+  if (typ == ReportTypeSignalUnsafe)
+    return "signal-unsafe call inside of a signal";
+  if (typ == ReportTypeErrnoInSignal)
+    return "signal handler spoils errno";
+  return "";
 }
 
 void PrintStack(const ReportStack *ent) {
@@ -89,11 +89,17 @@ static void PrintMutexSet(Vector<ReportMopMutex> const& mset) {
   }
 }
 
+static const char *MopDesc(bool first, bool write, bool atomic) {
+  return atomic ? (first ? (write ? "Atomic write" : "Atomic read")
+                : (write ? "Previous atomic write" : "Previous atomic read"))
+                : (first ? (write ? "Write" : "Read")
+                : (write ? "Previous write" : "Previous read"));
+}
+
 static void PrintMop(const ReportMop *mop, bool first) {
   char thrbuf[kThreadBufSize];
   Printf("  %s of size %d at %p by %s",
-      (first ? (mop->write ? "Write" : "Read")
-             : (mop->write ? "Previous write" : "Previous read")),
+      MopDesc(first, mop->write, mop->atomic),
       mop->size, (void*)mop->addr,
       thread_name(thrbuf, mop->tid));
   PrintMutexSet(mop->mset);
@@ -135,7 +141,7 @@ static void PrintThread(const ReportThread *rt) {
   if (rt->id == 0)  // Little sense in describing the main thread.
     return;
   Printf("  Thread T%d", rt->id);
-  if (rt->name)
+  if (rt->name && rt->name[0] != '\0')
     Printf(" '%s'", rt->name);
   char thrbuf[kThreadBufSize];
   Printf(" (tid=%zu, %s) created by %s",
@@ -152,9 +158,29 @@ static void PrintSleep(const ReportStack *s) {
   PrintStack(s);
 }
 
+static ReportStack *ChooseSummaryStack(const ReportDesc *rep) {
+  if (rep->mops.Size())
+    return rep->mops[0]->stack;
+  if (rep->stacks.Size())
+    return rep->stacks[0];
+  if (rep->mutexes.Size())
+    return rep->mutexes[0]->stack;
+  if (rep->threads.Size())
+    return rep->threads[0]->stack;
+  return 0;
+}
+
+ReportStack *SkipTsanInternalFrames(ReportStack *ent) {
+  while (FrameIsInternal(ent) && ent->next)
+    ent = ent->next;
+  return ent;
+}
+
 void PrintReport(const ReportDesc *rep) {
   Printf("==================\n");
-  PrintHeader(rep->typ);
+  const char *rep_typ_str = ReportTypeString(rep->typ);
+  Printf("WARNING: ThreadSanitizer: %s (pid=%d)\n", rep_typ_str,
+         (int)internal_getpid());
 
   for (uptr i = 0; i < rep->stacks.Size(); i++) {
     if (i)
@@ -177,6 +203,12 @@ void PrintReport(const ReportDesc *rep) {
   for (uptr i = 0; i < rep->threads.Size(); i++)
     PrintThread(rep->threads[i]);
 
+  if (rep->typ == ReportTypeThreadLeak && rep->count > 1)
+    Printf("  And %d more similar thread leaks.\n\n", rep->count - 1);
+
+  if (ReportStack *ent = SkipTsanInternalFrames(ChooseSummaryStack(rep)))
+    ReportErrorSummary(rep_typ_str, ent->file, ent->line, ent->func);
+
   Printf("==================\n");
 }
 
diff --git a/lib/tsan/rtl/tsan_report.h b/lib/tsan/rtl/tsan_report.h
index f6715d1aae9b9..b2ce0dd66a272 100644
--- a/lib/tsan/rtl/tsan_report.h
+++ b/lib/tsan/rtl/tsan_report.h
@@ -20,6 +20,7 @@ namespace __tsan {
 
 enum ReportType {
   ReportTypeRace,
+  ReportTypeVptrRace,
   ReportTypeUseAfterFree,
   ReportTypeThreadLeak,
   ReportTypeMutexDestroyLocked,
@@ -48,6 +49,7 @@ struct ReportMop {
   uptr addr;
   int size;
   bool write;
+  bool atomic;
   Vector<ReportMopMutex> mset;
   ReportStack *stack;
 
@@ -100,6 +102,7 @@ class ReportDesc {
   Vector<ReportMutex*> mutexes;
   Vector<ReportThread*> threads;
   ReportStack *sleep;
+  int count;
 
   ReportDesc();
   ~ReportDesc();
diff --git a/lib/tsan/rtl/tsan_rtl.cc b/lib/tsan/rtl/tsan_rtl.cc
index 493ed2055dfcc..5924858c84c52 100644
--- a/lib/tsan/rtl/tsan_rtl.cc
+++ b/lib/tsan/rtl/tsan_rtl.cc
@@ -23,6 +23,7 @@
 #include "tsan_rtl.h"
 #include "tsan_mman.h"
 #include "tsan_suppressions.h"
+#include "tsan_symbolize.h"
 
 volatile int __tsan_resumed = 0;
 
@@ -37,17 +38,40 @@ THREADLOCAL char cur_thread_placeholder[sizeof(ThreadState)] ALIGNED(64);
 #endif
 static char ctx_placeholder[sizeof(Context)] ALIGNED(64);
 
+// Can be overriden by a front-end.
+bool CPP_WEAK OnFinalize(bool failed) {
+  return failed;
+}
+
 static Context *ctx;
 Context *CTX() {
   return ctx;
 }
 
+static char thread_registry_placeholder[sizeof(ThreadRegistry)];
+
+static ThreadContextBase *CreateThreadContext(u32 tid) {
+  // Map thread trace when context is created.
+  MapThreadTrace(GetThreadTrace(tid), TraceSize() * sizeof(Event));
+  MapThreadTrace(GetThreadTraceHeader(tid), sizeof(Trace));
+  new(ThreadTrace(tid)) Trace();
+  void *mem = internal_alloc(MBlockThreadContex, sizeof(ThreadContext));
+  return new(mem) ThreadContext(tid);
+}
+
+#ifndef TSAN_GO
+static const u32 kThreadQuarantineSize = 16;
+#else
+static const u32 kThreadQuarantineSize = 64;
+#endif
+
 Context::Context()
   : initialized()
   , report_mtx(MutexTypeReport, StatMtxReport)
   , nreported()
   , nmissed_expected()
-  , thread_mtx(MutexTypeThreads, StatMtxThreads)
+  , thread_registry(new(thread_registry_placeholder) ThreadRegistry(
+      CreateThreadContext, kMaxTid, kThreadQuarantineSize))
   , racy_stacks(MBlockRacyStacks)
   , racy_addresses(MBlockRacyAddresses)
   , fired_suppressions(MBlockRacyAddresses) {
@@ -60,10 +84,12 @@ ThreadState::ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
   : fast_state(tid, epoch)
   // Do not touch these, rely on zero initialization,
   // they may be accessed before the ctor.
-  // , fast_ignore_reads()
-  // , fast_ignore_writes()
+  // , ignore_reads_and_writes()
   // , in_rtl()
   , shadow_stack_pos(&shadow_stack[0])
+#ifndef TSAN_GO
+  , jmp_bufs(MBlockJmpBuf)
+#endif
   , tid(tid)
   , unique_id(unique_id)
   , stk_addr(stk_addr)
@@ -72,94 +98,74 @@ ThreadState::ThreadState(Context *ctx, int tid, int unique_id, u64 epoch,
   , tls_size(tls_size) {
 }
 
-ThreadContext::ThreadContext(int tid)
-  : tid(tid)
-  , unique_id()
-  , os_id()
-  , user_id()
-  , thr()
-  , status(ThreadStatusInvalid)
-  , detached()
-  , reuse_count()
-  , epoch0()
-  , epoch1()
-  , dead_info()
-  , dead_next()
-  , name() {
-}
-
-static void WriteMemoryProfile(char *buf, uptr buf_size, int num) {
-  uptr shadow = GetShadowMemoryConsumption();
-
-  int nthread = 0;
-  int nlivethread = 0;
-  uptr threadmem = 0;
-  {
-    Lock l(&ctx->thread_mtx);
-    for (unsigned i = 0; i < kMaxTid; i++) {
-      ThreadContext *tctx = ctx->threads[i];
-      if (tctx == 0)
-        continue;
-      nthread += 1;
-      threadmem += sizeof(ThreadContext);
-      if (tctx->status != ThreadStatusRunning)
-        continue;
-      nlivethread += 1;
-      threadmem += sizeof(ThreadState);
-    }
-  }
-
-  uptr nsync = 0;
-  uptr syncmem = CTX()->synctab.GetMemoryConsumption(&nsync);
-
-  internal_snprintf(buf, buf_size, "%d: shadow=%zuMB"
-                                   " thread=%zuMB(total=%d/live=%d)"
-                                   " sync=%zuMB(cnt=%zu)\n",
-    num,
-    shadow >> 20,
-    threadmem >> 20, nthread, nlivethread,
-    syncmem >> 20, nsync);
+static void MemoryProfiler(Context *ctx, fd_t fd, int i) {
+  uptr n_threads;
+  uptr n_running_threads;
+  ctx->thread_registry->GetNumberOfThreads(&n_threads, &n_running_threads);
+  InternalScopedBuffer<char> buf(4096);
+  internal_snprintf(buf.data(), buf.size(), "%d: nthr=%d nlive=%d\n",
+      i, n_threads, n_running_threads);
+  internal_write(fd, buf.data(), internal_strlen(buf.data()));
+  WriteMemoryProfile(buf.data(), buf.size());
+  internal_write(fd, buf.data(), internal_strlen(buf.data()));
 }
 
-static void MemoryProfileThread(void *arg) {
+static void BackgroundThread(void *arg) {
   ScopedInRtl in_rtl;
-  fd_t fd = (fd_t)(uptr)arg;
+  Context *ctx = CTX();
+  const u64 kMs2Ns = 1000 * 1000;
+
+  fd_t mprof_fd = kInvalidFd;
+  if (flags()->profile_memory && flags()->profile_memory[0]) {
+    InternalScopedBuffer<char> filename(4096);
+    internal_snprintf(filename.data(), filename.size(), "%s.%d",
+        flags()->profile_memory, (int)internal_getpid());
+    uptr openrv = OpenFile(filename.data(), true);
+    if (internal_iserror(openrv)) {
+      Printf("ThreadSanitizer: failed to open memory profile file '%s'\n",
+          &filename[0]);
+    } else {
+      mprof_fd = openrv;
+    }
+  }
+
+  u64 last_flush = NanoTime();
   for (int i = 0; ; i++) {
-    InternalScopedBuffer<char> buf(4096);
-    WriteMemoryProfile(buf.data(), buf.size(), i);
-    internal_write(fd, buf.data(), internal_strlen(buf.data()));
     SleepForSeconds(1);
-  }
-}
+    u64 now = NanoTime();
+
+    // Flush memory if requested.
+    if (flags()->flush_memory_ms) {
+      if (last_flush + flags()->flush_memory_ms * kMs2Ns < now) {
+        FlushShadowMemory();
+        last_flush = NanoTime();
+      }
+    }
 
-static void InitializeMemoryProfile() {
-  if (flags()->profile_memory == 0 || flags()->profile_memory[0] == 0)
-    return;
-  InternalScopedBuffer<char> filename(4096);
-  internal_snprintf(filename.data(), filename.size(), "%s.%d",
-      flags()->profile_memory, GetPid());
-  fd_t fd = internal_open(filename.data(), true);
-  if (fd == kInvalidFd) {
-    Printf("Failed to open memory profile file '%s'\n", &filename[0]);
-    Die();
-  }
-  internal_start_thread(&MemoryProfileThread, (void*)(uptr)fd);
-}
+    // Write memory profile if requested.
+    if (mprof_fd != kInvalidFd)
+      MemoryProfiler(ctx, mprof_fd, i);
 
-static void MemoryFlushThread(void *arg) {
-  ScopedInRtl in_rtl;
-  for (int i = 0; ; i++) {
-    SleepForMillis(flags()->flush_memory_ms);
-    FlushShadowMemory();
+#ifndef TSAN_GO
+    // Flush symbolizer cache if requested.
+    if (flags()->flush_symbolizer_ms > 0) {
+      u64 last = atomic_load(&ctx->last_symbolize_time_ns,
+                             memory_order_relaxed);
+      if (last != 0 && last + flags()->flush_symbolizer_ms * kMs2Ns < now) {
+        Lock l(&ctx->report_mtx);
+        SpinMutexLock l2(&CommonSanitizerReportMutex);
+        SymbolizeFlush();
+        atomic_store(&ctx->last_symbolize_time_ns, 0, memory_order_relaxed);
+      }
+    }
+#endif
   }
 }
 
-static void InitializeMemoryFlush() {
-  if (flags()->flush_memory_ms == 0)
-    return;
-  if (flags()->flush_memory_ms < 100)
-    flags()->flush_memory_ms = 100;
-  internal_start_thread(&MemoryFlushThread, 0);
+void DontNeedShadowFor(uptr addr, uptr size) {
+  uptr shadow_beg = MemToShadow(addr);
+  uptr shadow_end = MemToShadow(addr + size);
+  FlushUnneededShadowMemory(shadow_beg, shadow_end - shadow_beg);
 }
 
 void MapShadow(uptr addr, uptr size) {
@@ -182,6 +188,7 @@ void Initialize(ThreadState *thr) {
   if (is_initialized)
     return;
   is_initialized = true;
+  SanitizerToolName = "ThreadSanitizer";
   // Install tool-specific callbacks in sanitizer_common.
   SetCheckFailedCallback(TsanCheckFailed);
 
@@ -197,9 +204,6 @@ void Initialize(ThreadState *thr) {
 #ifndef TSAN_GO
   InitializeShadowMemory();
 #endif
-  ctx->dead_list_size = 0;
-  ctx->dead_list_head = 0;
-  ctx->dead_list_tail = 0;
   InitializeFlags(&ctx->flags, env);
   // Setup correct file descriptor for error reports.
   if (internal_strcmp(flags()->log_path, "stdout") == 0)
@@ -220,26 +224,24 @@ void Initialize(ThreadState *thr) {
     }
   }
 #endif
-  InitializeMemoryProfile();
-  InitializeMemoryFlush();
+  internal_start_thread(&BackgroundThread, 0);
 
   if (ctx->flags.verbosity)
     Printf("***** Running under ThreadSanitizer v2 (pid %d) *****\n",
-               GetPid());
+           (int)internal_getpid());
 
   // Initialize thread 0.
-  ctx->thread_seq = 0;
   int tid = ThreadCreate(thr, 0, 0, true);
   CHECK_EQ(tid, 0);
-  ThreadStart(thr, tid, GetPid());
+  ThreadStart(thr, tid, internal_getpid());
   CHECK_EQ(thr->in_rtl, 1);
   ctx->initialized = true;
 
   if (flags()->stop_on_start) {
     Printf("ThreadSanitizer is suspended at startup (pid %d)."
            " Call __tsan_resume().\n",
-           GetPid());
-    while (__tsan_resumed == 0);
+           (int)internal_getpid());
+    while (__tsan_resumed == 0) {}
   }
 }
 
@@ -253,8 +255,15 @@ int Finalize(ThreadState *thr) {
 
   // Wait for pending reports.
   ctx->report_mtx.Lock();
+  CommonSanitizerReportMutex.Lock();
+  CommonSanitizerReportMutex.Unlock();
   ctx->report_mtx.Unlock();
 
+#ifndef TSAN_GO
+  if (ctx->flags.verbosity)
+    AllocatorPrintStats();
+#endif
+
   ThreadFinalize(thr);
 
   if (ctx->nreported) {
@@ -272,6 +281,15 @@ int Finalize(ThreadState *thr) {
         ctx->nmissed_expected);
   }
 
+  if (flags()->print_suppressions)
+    PrintMatchedSuppressions();
+#ifndef TSAN_GO
+  if (flags()->print_benign)
+    PrintMatchedBenignRaces();
+#endif
+
+  failed = OnFinalize(failed);
+
   StatAggregate(ctx->stat, thr->stat);
   StatOutput(ctx->stat);
   return failed ? flags()->exitcode : 0;
@@ -296,15 +314,20 @@ u32 CurrentStackId(ThreadState *thr, uptr pc) {
 void TraceSwitch(ThreadState *thr) {
   thr->nomalloc++;
   ScopedInRtl in_rtl;
-  Lock l(&thr->trace.mtx);
+  Trace *thr_trace = ThreadTrace(thr->tid);
+  Lock l(&thr_trace->mtx);
   unsigned trace = (thr->fast_state.epoch() / kTracePartSize) % TraceParts();
-  TraceHeader *hdr = &thr->trace.headers[trace];
+  TraceHeader *hdr = &thr_trace->headers[trace];
   hdr->epoch0 = thr->fast_state.epoch();
   hdr->stack0.ObtainCurrent(thr, 0);
   hdr->mset0 = thr->mset;
   thr->nomalloc--;
 }
 
+Trace *ThreadTrace(int tid) {
+  return (Trace*)GetThreadTraceHeader(tid);
+}
+
 uptr TraceTopPC(ThreadState *thr) {
   Event *events = (Event*)GetThreadTrace(thr->tid);
   uptr pc = events[thr->fast_state.GetTracePos()];
@@ -330,18 +353,18 @@ extern "C" void __tsan_report_race() {
 #endif
 
 ALWAYS_INLINE
-static Shadow LoadShadow(u64 *p) {
+Shadow LoadShadow(u64 *p) {
   u64 raw = atomic_load((atomic_uint64_t*)p, memory_order_relaxed);
   return Shadow(raw);
 }
 
 ALWAYS_INLINE
-static void StoreShadow(u64 *sp, u64 s) {
+void StoreShadow(u64 *sp, u64 s) {
   atomic_store((atomic_uint64_t*)sp, s, memory_order_relaxed);
 }
 
 ALWAYS_INLINE
-static void StoreIfNotYetStored(u64 *sp, u64 *s) {
+void StoreIfNotYetStored(u64 *sp, u64 *s) {
   StoreShadow(sp, *s);
   *s = 0;
 }
@@ -358,18 +381,6 @@ static inline void HandleRace(ThreadState *thr, u64 *shadow_mem,
 #endif
 }
 
-static inline bool BothReads(Shadow s, int kAccessIsWrite) {
-  return !kAccessIsWrite && !s.is_write();
-}
-
-static inline bool OldIsRWNotWeaker(Shadow old, int kAccessIsWrite) {
-  return old.is_write() || !kAccessIsWrite;
-}
-
-static inline bool OldIsRWWeakerOrEqual(Shadow old, int kAccessIsWrite) {
-  return !old.is_write() || kAccessIsWrite;
-}
-
 static inline bool OldIsInSameSynchEpoch(Shadow old, ThreadState *thr) {
   return old.epoch() >= thr->fast_synch_epoch;
 }
@@ -378,9 +389,9 @@ static inline bool HappensBefore(Shadow old, ThreadState *thr) {
   return thr->clock.get(old.TidWithIgnore()) >= old.epoch();
 }
 
-ALWAYS_INLINE
+ALWAYS_INLINE USED
 void MemoryAccessImpl(ThreadState *thr, uptr addr,
-    int kAccessSizeLog, bool kAccessIsWrite,
+    int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
     u64 *shadow_mem, Shadow cur) {
   StatInc(thr, StatMop);
   StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
@@ -452,9 +463,30 @@ void MemoryAccessImpl(ThreadState *thr, uptr addr,
   return;
 }
 
-ALWAYS_INLINE
+void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
+    int size, bool kAccessIsWrite, bool kIsAtomic) {
+  while (size) {
+    int size1 = 1;
+    int kAccessSizeLog = kSizeLog1;
+    if (size >= 8 && (addr & ~7) == ((addr + 8) & ~7)) {
+      size1 = 8;
+      kAccessSizeLog = kSizeLog8;
+    } else if (size >= 4 && (addr & ~7) == ((addr + 4) & ~7)) {
+      size1 = 4;
+      kAccessSizeLog = kSizeLog4;
+    } else if (size >= 2 && (addr & ~7) == ((addr + 2) & ~7)) {
+      size1 = 2;
+      kAccessSizeLog = kSizeLog2;
+    }
+    MemoryAccess(thr, pc, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic);
+    addr += size1;
+    size -= size1;
+  }
+}
+
+ALWAYS_INLINE USED
 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
-    int kAccessSizeLog, bool kAccessIsWrite) {
+    int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic) {
   u64 *shadow_mem = (u64*)MemToShadow(addr);
   DPrintf2("#%d: MemoryAccess: @%p %p size=%d"
       " is_write=%d shadow_mem=%p {%zx, %zx, %zx, %zx}\n",
@@ -473,6 +505,16 @@ void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
   }
 #endif
 
+  if (*shadow_mem == kShadowRodata) {
+    // Access to .rodata section, no races here.
+    // Measurements show that it can be 10-20% of all memory accesses.
+    StatInc(thr, StatMop);
+    StatInc(thr, kAccessIsWrite ? StatMopWrite : StatMopRead);
+    StatInc(thr, (StatType)(StatMop1 + kAccessSizeLog));
+    StatInc(thr, StatMopRodata);
+    return;
+  }
+
   FastState fast_state = thr->fast_state;
   if (fast_state.GetIgnoreBit())
     return;
@@ -481,17 +523,20 @@ void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
   Shadow cur(fast_state);
   cur.SetAddr0AndSizeLog(addr & 7, kAccessSizeLog);
   cur.SetWrite(kAccessIsWrite);
+  cur.SetAtomic(kIsAtomic);
 
   // We must not store to the trace if we do not store to the shadow.
   // That is, this call must be moved somewhere below.
   TraceAddEvent(thr, fast_state, EventTypeMop, pc);
 
-  MemoryAccessImpl(thr, addr, kAccessSizeLog, kAccessIsWrite,
+  MemoryAccessImpl(thr, addr, kAccessSizeLog, kAccessIsWrite, kIsAtomic,
       shadow_mem, cur);
 }
 
 static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
                            u64 val) {
+  (void)thr;
+  (void)pc;
   if (size == 0)
     return;
   // FIXME: fix me.
@@ -508,23 +553,42 @@ static void MemoryRangeSet(ThreadState *thr, uptr pc, uptr addr, uptr size,
   // let it just crash as usual.
   if (!IsAppMem(addr) || !IsAppMem(addr + size - 1))
     return;
-  (void)thr;
-  (void)pc;
-  // Some programs mmap like hundreds of GBs but actually used a small part.
-  // So, it's better to report a false positive on the memory
-  // then to hang here senselessly.
-  const uptr kMaxResetSize = 4ull*1024*1024*1024;
-  if (size > kMaxResetSize)
-    size = kMaxResetSize;
+  // Don't want to touch lots of shadow memory.
+  // If a program maps 10MB stack, there is no need reset the whole range.
   size = (size + (kShadowCell - 1)) & ~(kShadowCell - 1);
-  u64 *p = (u64*)MemToShadow(addr);
-  CHECK(IsShadowMem((uptr)p));
-  CHECK(IsShadowMem((uptr)(p + size * kShadowCnt / kShadowCell - 1)));
-  // FIXME: may overwrite a part outside the region
-  for (uptr i = 0; i < size * kShadowCnt / kShadowCell;) {
-    p[i++] = val;
-    for (uptr j = 1; j < kShadowCnt; j++)
-      p[i++] = 0;
+  if (size < 64*1024) {
+    u64 *p = (u64*)MemToShadow(addr);
+    CHECK(IsShadowMem((uptr)p));
+    CHECK(IsShadowMem((uptr)(p + size * kShadowCnt / kShadowCell - 1)));
+    // FIXME: may overwrite a part outside the region
+    for (uptr i = 0; i < size / kShadowCell * kShadowCnt;) {
+      p[i++] = val;
+      for (uptr j = 1; j < kShadowCnt; j++)
+        p[i++] = 0;
+    }
+  } else {
+    // The region is big, reset only beginning and end.
+    const uptr kPageSize = 4096;
+    u64 *begin = (u64*)MemToShadow(addr);
+    u64 *end = begin + size / kShadowCell * kShadowCnt;
+    u64 *p = begin;
+    // Set at least first kPageSize/2 to page boundary.
+    while ((p < begin + kPageSize / kShadowSize / 2) || ((uptr)p % kPageSize)) {
+      *p++ = val;
+      for (uptr j = 1; j < kShadowCnt; j++)
+        *p++ = 0;
+    }
+    // Reset middle part.
+    u64 *p1 = p;
+    p = RoundDown(end, kPageSize);
+    UnmapOrDie((void*)p1, (uptr)p - (uptr)p1);
+    MmapFixedNoReserve((uptr)p1, (uptr)p - (uptr)p1);
+    // Set the ending.
+    while (p < end) {
+      *p++ = val;
+      for (uptr j = 1; j < kShadowCnt; j++)
+        *p++ = 0;
+    }
   }
 }
 
@@ -533,7 +597,17 @@ void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 }
 
 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size) {
+  // Processing more than 1k (4k of shadow) is expensive,
+  // can cause excessive memory consumption (user does not necessary touch
+  // the whole range) and most likely unnecessary.
+  if (size > 1024)
+    size = 1024;
+  CHECK_EQ(thr->is_freeing, false);
+  thr->is_freeing = true;
   MemoryAccessRange(thr, pc, addr, size, true);
+  thr->is_freeing = false;
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
   Shadow s(thr->fast_state);
   s.ClearIgnoreBit();
   s.MarkAsFreed();
@@ -543,6 +617,8 @@ void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size) {
 }
 
 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size) {
+  thr->fast_state.IncrementEpoch();
+  TraceAddEvent(thr, thr->fast_state, EventTypeMop, pc);
   Shadow s(thr->fast_state);
   s.ClearIgnoreBit();
   s.SetWrite(true);
@@ -550,7 +626,7 @@ void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size) {
   MemoryRangeSet(thr, pc, addr, size, s.raw());
 }
 
-ALWAYS_INLINE
+ALWAYS_INLINE USED
 void FuncEntry(ThreadState *thr, uptr pc) {
   DCHECK_EQ(thr->in_rtl, 0);
   StatInc(thr, StatFuncEnter);
@@ -580,7 +656,7 @@ void FuncEntry(ThreadState *thr, uptr pc) {
   thr->shadow_stack_pos++;
 }
 
-ALWAYS_INLINE
+ALWAYS_INLINE USED
 void FuncExit(ThreadState *thr) {
   DCHECK_EQ(thr->in_rtl, 0);
   StatInc(thr, StatFuncExit);
diff --git a/lib/tsan/rtl/tsan_rtl.h b/lib/tsan/rtl/tsan_rtl.h
index 6b0ab0d385ef7..f1a73e457331c 100644
--- a/lib/tsan/rtl/tsan_rtl.h
+++ b/lib/tsan/rtl/tsan_rtl.h
@@ -26,8 +26,9 @@
 #ifndef TSAN_RTL_H
 #define TSAN_RTL_H
 
-#include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_allocator.h"
+#include "sanitizer_common/sanitizer_common.h"
+#include "sanitizer_common/sanitizer_thread_registry.h"
 #include "tsan_clock.h"
 #include "tsan_defs.h"
 #include "tsan_flags.h"
@@ -46,15 +47,73 @@ namespace __tsan {
 
 // Descriptor of user's memory block.
 struct MBlock {
-  Mutex mtx;
-  uptr size;
-  u32 alloc_tid;
-  u32 alloc_stack_id;
-  SyncVar *head;
+  /*
+  u64 mtx : 1;  // must be first
+  u64 lst : 44;
+  u64 stk : 31;  // on word boundary
+  u64 tid : kTidBits;
+  u64 siz : 128 - 1 - 31 - 44 - kTidBits;  // 39
+  */
+  u64 raw[2];
+
+  void Init(uptr siz, u32 tid, u32 stk) {
+    raw[0] = raw[1] = 0;
+    raw[1] |= (u64)siz << ((1 + 44 + 31 + kTidBits) % 64);
+    raw[1] |= (u64)tid << ((1 + 44 + 31) % 64);
+    raw[0] |= (u64)stk << (1 + 44);
+    raw[1] |= (u64)stk >> (64 - 44 - 1);
+    DCHECK_EQ(Size(), siz);
+    DCHECK_EQ(Tid(), tid);
+    DCHECK_EQ(StackId(), stk);
+  }
+
+  u32 Tid() const {
+    return GetLsb(raw[1] >> ((1 + 44 + 31) % 64), kTidBits);
+  }
+
+  uptr Size() const {
+    return raw[1] >> ((1 + 31 + 44 + kTidBits) % 64);
+  }
+
+  u32 StackId() const {
+    return (raw[0] >> (1 + 44)) | GetLsb(raw[1] << (64 - 44 - 1), 31);
+  }
+
+  SyncVar *ListHead() const {
+    return (SyncVar*)(GetLsb(raw[0] >> 1, 44) << 3);
+  }
+
+  void ListPush(SyncVar *v) {
+    SyncVar *lst = ListHead();
+    v->next = lst;
+    u64 x = (u64)v ^ (u64)lst;
+    x = (x >> 3) << 1;
+    raw[0] ^= x;
+    DCHECK_EQ(ListHead(), v);
+  }
+
+  SyncVar *ListPop() {
+    SyncVar *lst = ListHead();
+    SyncVar *nxt = lst->next;
+    lst->next = 0;
+    u64 x = (u64)lst ^ (u64)nxt;
+    x = (x >> 3) << 1;
+    raw[0] ^= x;
+    DCHECK_EQ(ListHead(), nxt);
+    return lst;
+  }
 
-  MBlock()
-    : mtx(MutexTypeMBlock, StatMtxMBlock) {
+  void ListReset() {
+    SyncVar *lst = ListHead();
+    u64 x = (u64)lst;
+    x = (x >> 3) << 1;
+    raw[0] ^= x;
+    DCHECK_EQ(ListHead(), 0);
   }
+
+  void Lock();
+  void Unlock();
+  typedef GenericScopedLock<MBlock> ScopedLock;
 };
 
 #ifndef TSAN_GO
@@ -65,22 +124,11 @@ const uptr kAllocatorSpace = 0x7d0000000000ULL;
 #endif
 const uptr kAllocatorSize  =  0x10000000000ULL;  // 1T.
 
-struct TsanMapUnmapCallback {
-  void OnMap(uptr p, uptr size) const { }
-  void OnUnmap(uptr p, uptr size) const {
-    // We are about to unmap a chunk of user memory.
-    // Mark the corresponding shadow memory as not needed.
-    uptr shadow_beg = MemToShadow(p);
-    uptr shadow_end = MemToShadow(p + size);
-    CHECK(IsAligned(shadow_end|shadow_beg, GetPageSizeCached()));
-    FlushUnneededShadowMemory(shadow_beg, shadow_end - shadow_beg);
-  }
-};
-
+struct MapUnmapCallback;
 typedef SizeClassAllocator64<kAllocatorSpace, kAllocatorSize, sizeof(MBlock),
-    DefaultSizeClassMap> PrimaryAllocator;
+    DefaultSizeClassMap, MapUnmapCallback> PrimaryAllocator;
 typedef SizeClassAllocatorLocalCache<PrimaryAllocator> AllocatorCache;
-typedef LargeMmapAllocator<TsanMapUnmapCallback> SecondaryAllocator;
+typedef LargeMmapAllocator<MapUnmapCallback> SecondaryAllocator;
 typedef CombinedAllocator<PrimaryAllocator, AllocatorCache,
     SecondaryAllocator> Allocator;
 Allocator *allocator();
@@ -89,6 +137,8 @@ Allocator *allocator();
 void TsanCheckFailed(const char *file, int line, const char *cond,
                      u64 v1, u64 v2);
 
+const u64 kShadowRodata = (u64)-1;  // .rodata shadow marker
+
 // FastState (from most significant bit):
 //   ignore          : 1
 //   tid             : kTidBits
@@ -173,7 +223,8 @@ class FastState {
 //   freed           : 1
 //   tid             : kTidBits
 //   epoch           : kClkBits
-//   is_write        : 1
+//   is_atomic       : 1
+//   is_read         : 1
 //   size_log        : 2
 //   addr0           : 3
 class Shadow : public FastState {
@@ -197,13 +248,26 @@ class Shadow : public FastState {
   }
 
   void SetWrite(unsigned kAccessIsWrite) {
-    DCHECK_EQ(x_ & 32, 0);
-    if (kAccessIsWrite)
-      x_ |= 32;
-    DCHECK_EQ(kAccessIsWrite, is_write());
+    DCHECK_EQ(x_ & kReadBit, 0);
+    if (!kAccessIsWrite)
+      x_ |= kReadBit;
+    DCHECK_EQ(kAccessIsWrite, IsWrite());
+  }
+
+  void SetAtomic(bool kIsAtomic) {
+    DCHECK(!IsAtomic());
+    if (kIsAtomic)
+      x_ |= kAtomicBit;
+    DCHECK_EQ(IsAtomic(), kIsAtomic);
   }
 
-  bool IsZero() const { return x_ == 0; }
+  bool IsAtomic() const {
+    return x_ & kAtomicBit;
+  }
+
+  bool IsZero() const {
+    return x_ == 0;
+  }
 
   static inline bool TidsAreEqual(const Shadow s1, const Shadow s2) {
     u64 shifted_xor = (s1.x_ ^ s2.x_) >> kTidShift;
@@ -250,7 +314,8 @@ class Shadow : public FastState {
   }
   u64 addr0() const { return x_ & 7; }
   u64 size() const { return 1ull << size_log(); }
-  bool is_write() const { return x_ & 32; }
+  bool IsWrite() const { return !IsRead(); }
+  bool IsRead() const { return x_ & kReadBit; }
 
   // The idea behind the freed bit is as follows.
   // When the memory is freed (or otherwise unaccessible) we write to the shadow
@@ -265,13 +330,46 @@ class Shadow : public FastState {
      x_ |= kFreedBit;
   }
 
+  bool IsFreed() const {
+    return x_ & kFreedBit;
+  }
+
   bool GetFreedAndReset() {
     bool res = x_ & kFreedBit;
     x_ &= ~kFreedBit;
     return res;
   }
 
+  bool IsBothReadsOrAtomic(bool kIsWrite, bool kIsAtomic) const {
+    // analyzes 5-th bit (is_read) and 6-th bit (is_atomic)
+    bool v = x_ & u64(((kIsWrite ^ 1) << kReadShift)
+        | (kIsAtomic << kAtomicShift));
+    DCHECK_EQ(v, (!IsWrite() && !kIsWrite) || (IsAtomic() && kIsAtomic));
+    return v;
+  }
+
+  bool IsRWNotWeaker(bool kIsWrite, bool kIsAtomic) const {
+    bool v = ((x_ >> kReadShift) & 3)
+        <= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
+    DCHECK_EQ(v, (IsAtomic() < kIsAtomic) ||
+        (IsAtomic() == kIsAtomic && !IsWrite() <= !kIsWrite));
+    return v;
+  }
+
+  bool IsRWWeakerOrEqual(bool kIsWrite, bool kIsAtomic) const {
+    bool v = ((x_ >> kReadShift) & 3)
+        >= u64((kIsWrite ^ 1) | (kIsAtomic << 1));
+    DCHECK_EQ(v, (IsAtomic() > kIsAtomic) ||
+        (IsAtomic() == kIsAtomic && !IsWrite() >= !kIsWrite));
+    return v;
+  }
+
  private:
+  static const u64 kReadShift   = 5;
+  static const u64 kReadBit     = 1ull << kReadShift;
+  static const u64 kAtomicShift = 6;
+  static const u64 kAtomicBit   = 1ull << kAtomicShift;
+
   u64 size_log() const { return (x_ >> 3) & 3; }
 
   static bool TwoRangesIntersectSLOW(const Shadow s1, const Shadow s2) {
@@ -286,6 +384,12 @@ class Shadow : public FastState {
 
 struct SignalContext;
 
+struct JmpBuf {
+  uptr sp;
+  uptr mangled_sp;
+  uptr *shadow_stack_pos;
+};
+
 // This struct is stored in TLS.
 struct ThreadState {
   FastState fast_state;
@@ -308,7 +412,6 @@ struct ThreadState {
   uptr *shadow_stack_pos;
   u64 *racy_shadow_addr;
   u64 racy_state[2];
-  Trace trace;
 #ifndef TSAN_GO
   // C/C++ uses embed shadow stack of fixed size.
   uptr shadow_stack[kShadowStackSize];
@@ -321,12 +424,16 @@ struct ThreadState {
   ThreadClock clock;
 #ifndef TSAN_GO
   AllocatorCache alloc_cache;
+  Vector<JmpBuf> jmp_bufs;
 #endif
   u64 stat[StatCnt];
   const int tid;
   const int unique_id;
   int in_rtl;
+  bool in_symbolizer;
   bool is_alive;
+  bool is_freeing;
+  bool is_vptr_access;
   const uptr stk_addr;
   const uptr stk_size;
   const uptr tls_addr;
@@ -360,41 +467,30 @@ INLINE ThreadState *cur_thread() {
 }
 #endif
 
-enum ThreadStatus {
-  ThreadStatusInvalid,   // Non-existent thread, data is invalid.
-  ThreadStatusCreated,   // Created but not yet running.
-  ThreadStatusRunning,   // The thread is currently running.
-  ThreadStatusFinished,  // Joinable thread is finished but not yet joined.
-  ThreadStatusDead       // Joined, but some info (trace) is still alive.
-};
-
-// An info about a thread that is hold for some time after its termination.
-struct ThreadDeadInfo {
-  Trace trace;
-};
-
-struct ThreadContext {
-  const int tid;
-  int unique_id;  // Non-rolling thread id.
-  uptr os_id;  // pid
-  uptr user_id;  // Some opaque user thread id (e.g. pthread_t).
+class ThreadContext : public ThreadContextBase {
+ public:
+  explicit ThreadContext(int tid);
+  ~ThreadContext();
   ThreadState *thr;
-  ThreadStatus status;
-  bool detached;
-  int reuse_count;
+#ifdef TSAN_GO
+  StackTrace creation_stack;
+#else
+  u32 creation_stack_id;
+#endif
   SyncClock sync;
   // Epoch at which the thread had started.
   // If we see an event from the thread stamped by an older epoch,
   // the event is from a dead thread that shared tid with this thread.
   u64 epoch0;
   u64 epoch1;
-  StackTrace creation_stack;
-  int creation_tid;
-  ThreadDeadInfo *dead_info;
-  ThreadContext *dead_next;  // In dead thread list.
-  char *name;  // As annotated by user.
 
-  explicit ThreadContext(int tid);
+  // Override superclass callbacks.
+  void OnDead();
+  void OnJoined(void *arg);
+  void OnFinished();
+  void OnStarted(void *arg);
+  void OnCreated(void *arg);
+  void OnReset();
 };
 
 struct RacyStacks {
@@ -416,6 +512,7 @@ struct RacyAddress {
 struct FiredSuppression {
   ReportType type;
   uptr pc;
+  Suppression *supp;
 };
 
 struct Context {
@@ -428,16 +525,9 @@ struct Context {
   Mutex report_mtx;
   int nreported;
   int nmissed_expected;
+  atomic_uint64_t last_symbolize_time_ns;
 
-  Mutex thread_mtx;
-  unsigned thread_seq;
-  unsigned unique_thread_seq;
-  int alive_threads;
-  int max_alive_threads;
-  ThreadContext *threads[kMaxTid];
-  int dead_list_size;
-  ThreadContext* dead_list_head;
-  ThreadContext* dead_list_tail;
+  ThreadRegistry *thread_registry;
 
   Vector<RacyStacks> racy_stacks;
   Vector<RacyAddress> racy_addresses;
@@ -472,6 +562,7 @@ class ScopedReport {
   void AddMutex(const SyncVar *s);
   void AddLocation(uptr addr, uptr size);
   void AddSleep(u32 stack_id);
+  void SetCount(int count);
 
   const ReportDesc *GetReport() const;
 
@@ -489,13 +580,18 @@ void RestoreStack(int tid, const u64 epoch, StackTrace *stk, MutexSet *mset);
 
 void StatAggregate(u64 *dst, u64 *src);
 void StatOutput(u64 *stat);
-void ALWAYS_INLINE INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
+void ALWAYS_INLINE StatInc(ThreadState *thr, StatType typ, u64 n = 1) {
   if (kCollectStats)
     thr->stat[typ] += n;
 }
+void ALWAYS_INLINE StatSet(ThreadState *thr, StatType typ, u64 n) {
+  if (kCollectStats)
+    thr->stat[typ] = n;
+}
 
 void MapShadow(uptr addr, uptr size);
 void MapThreadTrace(uptr addr, uptr size);
+void DontNeedShadowFor(uptr addr, uptr size);
 void InitializeShadowMemory();
 void InitializeInterceptors();
 void InitializeDynamicAnnotations();
@@ -503,11 +599,15 @@ void InitializeDynamicAnnotations();
 void ReportRace(ThreadState *thr);
 bool OutputReport(Context *ctx,
                   const ScopedReport &srep,
-                  const ReportStack *suppress_stack = 0);
+                  const ReportStack *suppress_stack1 = 0,
+                  const ReportStack *suppress_stack2 = 0);
 bool IsFiredSuppression(Context *ctx,
                         const ScopedReport &srep,
                         const StackTrace &trace);
 bool IsExpectedReport(uptr addr, uptr size);
+void PrintMatchedBenignRaces();
+bool FrameIsInternal(const ReportStack *frame);
+ReportStack *SkipTsanInternalFrames(ReportStack *ent);
 
 #if defined(TSAN_DEBUG_OUTPUT) && TSAN_DEBUG_OUTPUT >= 1
 # define DPrintf Printf
@@ -523,6 +623,7 @@ bool IsExpectedReport(uptr addr, uptr size);
 
 u32 CurrentStackId(ThreadState *thr, uptr pc);
 void PrintCurrentStack(ThreadState *thr, uptr pc);
+void PrintCurrentStackSlow();  // uses libunwind
 
 void Initialize(ThreadState *thr);
 int Finalize(ThreadState *thr);
@@ -532,16 +633,42 @@ SyncVar* GetJavaSync(ThreadState *thr, uptr pc, uptr addr,
 SyncVar* GetAndRemoveJavaSync(ThreadState *thr, uptr pc, uptr addr);
 
 void MemoryAccess(ThreadState *thr, uptr pc, uptr addr,
-    int kAccessSizeLog, bool kAccessIsWrite);
+    int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic);
 void MemoryAccessImpl(ThreadState *thr, uptr addr,
-    int kAccessSizeLog, bool kAccessIsWrite,
+    int kAccessSizeLog, bool kAccessIsWrite, bool kIsAtomic,
     u64 *shadow_mem, Shadow cur);
-void MemoryRead1Byte(ThreadState *thr, uptr pc, uptr addr);
-void MemoryWrite1Byte(ThreadState *thr, uptr pc, uptr addr);
-void MemoryRead8Byte(ThreadState *thr, uptr pc, uptr addr);
-void MemoryWrite8Byte(ThreadState *thr, uptr pc, uptr addr);
 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
-                       uptr size, bool is_write);
+    uptr size, bool is_write);
+void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
+    uptr size, uptr step, bool is_write);
+void UnalignedMemoryAccess(ThreadState *thr, uptr pc, uptr addr,
+    int size, bool kAccessIsWrite, bool kIsAtomic);
+
+const int kSizeLog1 = 0;
+const int kSizeLog2 = 1;
+const int kSizeLog4 = 2;
+const int kSizeLog8 = 3;
+
+void ALWAYS_INLINE MemoryRead(ThreadState *thr, uptr pc,
+                                     uptr addr, int kAccessSizeLog) {
+  MemoryAccess(thr, pc, addr, kAccessSizeLog, false, false);
+}
+
+void ALWAYS_INLINE MemoryWrite(ThreadState *thr, uptr pc,
+                                      uptr addr, int kAccessSizeLog) {
+  MemoryAccess(thr, pc, addr, kAccessSizeLog, true, false);
+}
+
+void ALWAYS_INLINE MemoryReadAtomic(ThreadState *thr, uptr pc,
+                                           uptr addr, int kAccessSizeLog) {
+  MemoryAccess(thr, pc, addr, kAccessSizeLog, false, true);
+}
+
+void ALWAYS_INLINE MemoryWriteAtomic(ThreadState *thr, uptr pc,
+                                            uptr addr, int kAccessSizeLog) {
+  MemoryAccess(thr, pc, addr, kAccessSizeLog, true, true);
+}
+
 void MemoryResetRange(ThreadState *thr, uptr pc, uptr addr, uptr size);
 void MemoryRangeFreed(ThreadState *thr, uptr pc, uptr addr, uptr size);
 void MemoryRangeImitateWrite(ThreadState *thr, uptr pc, uptr addr, uptr size);
@@ -564,8 +691,8 @@ void ProcessPendingSignals(ThreadState *thr);
 void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
                  bool rw, bool recursive, bool linker_init);
 void MutexDestroy(ThreadState *thr, uptr pc, uptr addr);
-void MutexLock(ThreadState *thr, uptr pc, uptr addr);
-void MutexUnlock(ThreadState *thr, uptr pc, uptr addr);
+void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec = 1);
+int  MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all = false);
 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr);
 void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr);
 void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr);
@@ -601,9 +728,10 @@ void TraceSwitch(ThreadState *thr);
 uptr TraceTopPC(ThreadState *thr);
 uptr TraceSize();
 uptr TraceParts();
+Trace *ThreadTrace(int tid);
 
 extern "C" void __tsan_trace_switch();
-void ALWAYS_INLINE INLINE TraceAddEvent(ThreadState *thr, FastState fs,
+void ALWAYS_INLINE TraceAddEvent(ThreadState *thr, FastState fs,
                                         EventType typ, u64 addr) {
   DCHECK_GE((int)typ, 0);
   DCHECK_LE((int)typ, 7);
diff --git a/lib/tsan/rtl/tsan_rtl_amd64.S b/lib/tsan/rtl/tsan_rtl_amd64.S
index af878563573ef..11c75c72dbe5b 100644
--- a/lib/tsan/rtl/tsan_rtl_amd64.S
+++ b/lib/tsan/rtl/tsan_rtl_amd64.S
@@ -160,6 +160,143 @@ __tsan_report_race_thunk:
   ret
   .cfi_endproc
 
+.hidden __tsan_setjmp
+.comm _ZN14__interception11real_setjmpE,8,8
+.globl setjmp
+.type setjmp, @function
+setjmp:
+  .cfi_startproc
+  // save env parameter
+  push %rdi
+  .cfi_adjust_cfa_offset 8
+  .cfi_rel_offset %rdi, 0
+  // obtain %rsp
+  lea 16(%rsp), %rdi
+  mov %rdi, %rsi
+  xor %fs:0x30, %rsi  // magic mangling of rsp (see libc setjmp)
+  rol $0x11, %rsi
+  // call tsan interceptor
+  call __tsan_setjmp
+  // restore env parameter
+  pop %rdi
+  .cfi_adjust_cfa_offset -8
+  .cfi_restore %rdi
+  // tail jump to libc setjmp
+  movl $0, %eax
+  movq _ZN14__interception11real_setjmpE@GOTPCREL(%rip), %rdx
+  jmp *(%rdx)
+  .cfi_endproc
+.size setjmp, .-setjmp
+
+.comm _ZN14__interception12real__setjmpE,8,8
+.globl _setjmp
+.type _setjmp, @function
+_setjmp:
+  .cfi_startproc
+  // save env parameter
+  push %rdi
+  .cfi_adjust_cfa_offset 8
+  .cfi_rel_offset %rdi, 0
+  // obtain %rsp
+  lea 16(%rsp), %rdi
+  mov %rdi, %rsi
+  xor %fs:0x30, %rsi  // magic mangling of rsp (see libc setjmp)
+  rol $0x11, %rsi
+  // call tsan interceptor
+  call __tsan_setjmp
+  // restore env parameter
+  pop %rdi
+  .cfi_adjust_cfa_offset -8
+  .cfi_restore %rdi
+  // tail jump to libc setjmp
+  movl $0, %eax
+  movq _ZN14__interception12real__setjmpE@GOTPCREL(%rip), %rdx
+  jmp *(%rdx)
+  .cfi_endproc
+.size _setjmp, .-_setjmp
+
+.comm _ZN14__interception14real_sigsetjmpE,8,8
+.globl sigsetjmp
+.type sigsetjmp, @function
+sigsetjmp:
+  .cfi_startproc
+  // save env parameter
+  push %rdi
+  .cfi_adjust_cfa_offset 8
+  .cfi_rel_offset %rdi, 0
+  // save savesigs parameter
+  push %rsi
+  .cfi_adjust_cfa_offset 8
+  .cfi_rel_offset %rsi, 0
+  // align stack frame
+  sub $8, %rsp
+  .cfi_adjust_cfa_offset 8
+  // obtain %rsp
+  lea 32(%rsp), %rdi
+  mov %rdi, %rsi
+  xor %fs:0x30, %rsi  // magic mangling of rsp (see libc setjmp)
+  rol $0x11, %rsi
+  // call tsan interceptor
+  call __tsan_setjmp
+  // unalign stack frame
+  add $8, %rsp
+  .cfi_adjust_cfa_offset -8
+  // restore savesigs parameter
+  pop %rsi
+  .cfi_adjust_cfa_offset -8
+  .cfi_restore %rsi
+  // restore env parameter
+  pop %rdi
+  .cfi_adjust_cfa_offset -8
+  .cfi_restore %rdi
+  // tail jump to libc sigsetjmp
+  movl $0, %eax
+  movq _ZN14__interception14real_sigsetjmpE@GOTPCREL(%rip), %rdx
+  jmp *(%rdx)
+  .cfi_endproc
+.size sigsetjmp, .-sigsetjmp
+
+.comm _ZN14__interception16real___sigsetjmpE,8,8
+.globl __sigsetjmp
+.type __sigsetjmp, @function
+__sigsetjmp:
+  .cfi_startproc
+  // save env parameter
+  push %rdi
+  .cfi_adjust_cfa_offset 8
+  .cfi_rel_offset %rdi, 0
+  // save savesigs parameter
+  push %rsi
+  .cfi_adjust_cfa_offset 8
+  .cfi_rel_offset %rsi, 0
+  // align stack frame
+  sub $8, %rsp
+  .cfi_adjust_cfa_offset 8
+  // obtain %rsp
+  lea 32(%rsp), %rdi
+  mov %rdi, %rsi
+  xor %fs:0x30, %rsi  // magic mangling of rsp (see libc setjmp)
+  rol $0x11, %rsi
+  // call tsan interceptor
+  call __tsan_setjmp
+  // unalign stack frame
+  add $8, %rsp
+  .cfi_adjust_cfa_offset -8
+  // restore savesigs parameter
+  pop %rsi
+  .cfi_adjust_cfa_offset -8
+  .cfi_restore %rsi
+  // restore env parameter
+  pop %rdi
+  .cfi_adjust_cfa_offset -8
+  .cfi_restore %rdi
+  // tail jump to libc sigsetjmp
+  movl $0, %eax
+  movq _ZN14__interception16real___sigsetjmpE@GOTPCREL(%rip), %rdx
+  jmp *(%rdx)
+  .cfi_endproc
+.size __sigsetjmp, .-__sigsetjmp
+
 #ifdef __linux__
 /* We do not need executable stack.  */
 .section        .note.GNU-stack,"",@progbits
diff --git a/lib/tsan/rtl/tsan_rtl_mutex.cc b/lib/tsan/rtl/tsan_rtl_mutex.cc
index d812f12be560a..cf2e44dd09eec 100644
--- a/lib/tsan/rtl/tsan_rtl_mutex.cc
+++ b/lib/tsan/rtl/tsan_rtl_mutex.cc
@@ -26,8 +26,12 @@ void MutexCreate(ThreadState *thr, uptr pc, uptr addr,
   CHECK_GT(thr->in_rtl, 0);
   DPrintf("#%d: MutexCreate %zx\n", thr->tid, addr);
   StatInc(thr, StatMutexCreate);
-  if (!linker_init && IsAppMem(addr))
-    MemoryWrite1Byte(thr, pc, addr);
+  if (!linker_init && IsAppMem(addr)) {
+    CHECK(!thr->is_freeing);
+    thr->is_freeing = true;
+    MemoryWrite(thr, pc, addr, kSizeLog1);
+    thr->is_freeing = false;
+  }
   SyncVar *s = ctx->synctab.GetOrCreateAndLock(thr, pc, addr, true);
   s->is_rw = rw;
   s->is_recursive = recursive;
@@ -49,13 +53,17 @@ void MutexDestroy(ThreadState *thr, uptr pc, uptr addr) {
   SyncVar *s = ctx->synctab.GetAndRemove(thr, pc, addr);
   if (s == 0)
     return;
-  if (IsAppMem(addr))
-    MemoryWrite1Byte(thr, pc, addr);
+  if (IsAppMem(addr)) {
+    CHECK(!thr->is_freeing);
+    thr->is_freeing = true;
+    MemoryWrite(thr, pc, addr, kSizeLog1);
+    thr->is_freeing = false;
+  }
   if (flags()->report_destroy_locked
       && s->owner_tid != SyncVar::kInvalidTid
       && !s->is_broken) {
     s->is_broken = true;
-    Lock l(&ctx->thread_mtx);
+    ThreadRegistryLock l(ctx->thread_registry);
     ScopedReport rep(ReportTypeMutexDestroyLocked);
     rep.AddMutex(s);
     StackTrace trace;
@@ -71,11 +79,12 @@ void MutexDestroy(ThreadState *thr, uptr pc, uptr addr) {
   DestroyAndFree(s);
 }
 
-void MutexLock(ThreadState *thr, uptr pc, uptr addr) {
+void MutexLock(ThreadState *thr, uptr pc, uptr addr, int rec) {
   CHECK_GT(thr->in_rtl, 0);
-  DPrintf("#%d: MutexLock %zx\n", thr->tid, addr);
+  DPrintf("#%d: MutexLock %zx rec=%d\n", thr->tid, addr, rec);
+  CHECK_GT(rec, 0);
   if (IsAppMem(addr))
-    MemoryRead1Byte(thr, pc, addr);
+    MemoryReadAtomic(thr, pc, addr, kSizeLog1);
   SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, addr, true);
   thr->fast_state.IncrementEpoch();
   TraceAddEvent(thr, thr->fast_state, EventTypeLock, s->GetId());
@@ -99,19 +108,20 @@ void MutexLock(ThreadState *thr, uptr pc, uptr addr) {
   } else if (!s->is_recursive) {
     StatInc(thr, StatMutexRecLock);
   }
-  s->recursion++;
+  s->recursion += rec;
   thr->mset.Add(s->GetId(), true, thr->fast_state.epoch());
   s->mtx.Unlock();
 }
 
-void MutexUnlock(ThreadState *thr, uptr pc, uptr addr) {
+int MutexUnlock(ThreadState *thr, uptr pc, uptr addr, bool all) {
   CHECK_GT(thr->in_rtl, 0);
-  DPrintf("#%d: MutexUnlock %zx\n", thr->tid, addr);
+  DPrintf("#%d: MutexUnlock %zx all=%d\n", thr->tid, addr, all);
   if (IsAppMem(addr))
-    MemoryRead1Byte(thr, pc, addr);
+    MemoryReadAtomic(thr, pc, addr, kSizeLog1);
   SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, addr, true);
   thr->fast_state.IncrementEpoch();
   TraceAddEvent(thr, thr->fast_state, EventTypeUnlock, s->GetId());
+  int rec = 0;
   if (s->recursion == 0) {
     if (!s->is_broken) {
       s->is_broken = true;
@@ -125,7 +135,8 @@ void MutexUnlock(ThreadState *thr, uptr pc, uptr addr) {
       PrintCurrentStack(thr, pc);
     }
   } else {
-    s->recursion--;
+    rec = all ? s->recursion : 1;
+    s->recursion -= rec;
     if (s->recursion == 0) {
       StatInc(thr, StatMutexUnlock);
       s->owner_tid = SyncVar::kInvalidTid;
@@ -139,6 +150,7 @@ void MutexUnlock(ThreadState *thr, uptr pc, uptr addr) {
   }
   thr->mset.Del(s->GetId(), true);
   s->mtx.Unlock();
+  return rec;
 }
 
 void MutexReadLock(ThreadState *thr, uptr pc, uptr addr) {
@@ -146,7 +158,7 @@ void MutexReadLock(ThreadState *thr, uptr pc, uptr addr) {
   DPrintf("#%d: MutexReadLock %zx\n", thr->tid, addr);
   StatInc(thr, StatMutexReadLock);
   if (IsAppMem(addr))
-    MemoryRead1Byte(thr, pc, addr);
+    MemoryReadAtomic(thr, pc, addr, kSizeLog1);
   SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, addr, false);
   thr->fast_state.IncrementEpoch();
   TraceAddEvent(thr, thr->fast_state, EventTypeRLock, s->GetId());
@@ -167,7 +179,7 @@ void MutexReadUnlock(ThreadState *thr, uptr pc, uptr addr) {
   DPrintf("#%d: MutexReadUnlock %zx\n", thr->tid, addr);
   StatInc(thr, StatMutexReadUnlock);
   if (IsAppMem(addr))
-    MemoryRead1Byte(thr, pc, addr);
+    MemoryReadAtomic(thr, pc, addr, kSizeLog1);
   SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, addr, true);
   thr->fast_state.IncrementEpoch();
   TraceAddEvent(thr, thr->fast_state, EventTypeRUnlock, s->GetId());
@@ -188,7 +200,7 @@ void MutexReadOrWriteUnlock(ThreadState *thr, uptr pc, uptr addr) {
   CHECK_GT(thr->in_rtl, 0);
   DPrintf("#%d: MutexReadOrWriteUnlock %zx\n", thr->tid, addr);
   if (IsAppMem(addr))
-    MemoryRead1Byte(thr, pc, addr);
+    MemoryReadAtomic(thr, pc, addr, kSizeLog1);
   SyncVar *s = CTX()->synctab.GetOrCreateAndLock(thr, pc, addr, true);
   bool write = true;
   if (s->owner_tid == SyncVar::kInvalidTid) {
@@ -240,18 +252,19 @@ void Acquire(ThreadState *thr, uptr pc, uptr addr) {
   s->mtx.ReadUnlock();
 }
 
+static void UpdateClockCallback(ThreadContextBase *tctx_base, void *arg) {
+  ThreadState *thr = reinterpret_cast<ThreadState*>(arg);
+  ThreadContext *tctx = static_cast<ThreadContext*>(tctx_base);
+  if (tctx->status == ThreadStatusRunning)
+    thr->clock.set(tctx->tid, tctx->thr->fast_state.epoch());
+  else
+    thr->clock.set(tctx->tid, tctx->epoch1);
+}
+
 void AcquireGlobal(ThreadState *thr, uptr pc) {
-  Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  for (unsigned i = 0; i < kMaxTid; i++) {
-    ThreadContext *tctx = ctx->threads[i];
-    if (tctx == 0)
-      continue;
-    if (tctx->status == ThreadStatusRunning)
-      thr->clock.set(i, tctx->thr->fast_state.epoch());
-    else
-      thr->clock.set(i, tctx->epoch1);
-  }
+  ThreadRegistryLock l(CTX()->thread_registry);
+  CTX()->thread_registry->RunCallbackForEachThreadLocked(
+      UpdateClockCallback, thr);
 }
 
 void Release(ThreadState *thr, uptr pc, uptr addr) {
@@ -275,19 +288,20 @@ void ReleaseStore(ThreadState *thr, uptr pc, uptr addr) {
 }
 
 #ifndef TSAN_GO
+static void UpdateSleepClockCallback(ThreadContextBase *tctx_base, void *arg) {
+  ThreadState *thr = reinterpret_cast<ThreadState*>(arg);
+  ThreadContext *tctx = static_cast<ThreadContext*>(tctx_base);
+  if (tctx->status == ThreadStatusRunning)
+    thr->last_sleep_clock.set(tctx->tid, tctx->thr->fast_state.epoch());
+  else
+    thr->last_sleep_clock.set(tctx->tid, tctx->epoch1);
+}
+
 void AfterSleep(ThreadState *thr, uptr pc) {
-  Context *ctx = CTX();
   thr->last_sleep_stack_id = CurrentStackId(thr, pc);
-  Lock l(&ctx->thread_mtx);
-  for (unsigned i = 0; i < kMaxTid; i++) {
-    ThreadContext *tctx = ctx->threads[i];
-    if (tctx == 0)
-      continue;
-    if (tctx->status == ThreadStatusRunning)
-      thr->last_sleep_clock.set(i, tctx->thr->fast_state.epoch());
-    else
-      thr->last_sleep_clock.set(i, tctx->epoch1);
-  }
+  ThreadRegistryLock l(CTX()->thread_registry);
+  CTX()->thread_registry->RunCallbackForEachThreadLocked(
+      UpdateSleepClockCallback, thr);
 }
 #endif
 
diff --git a/lib/tsan/rtl/tsan_rtl_report.cc b/lib/tsan/rtl/tsan_rtl_report.cc
index 1a780e4b8070a..f77a7a2efa964 100644
--- a/lib/tsan/rtl/tsan_rtl_report.cc
+++ b/lib/tsan/rtl/tsan_rtl_report.cc
@@ -15,6 +15,7 @@
 #include "sanitizer_common/sanitizer_placement_new.h"
 #include "sanitizer_common/sanitizer_stackdepot.h"
 #include "sanitizer_common/sanitizer_common.h"
+#include "sanitizer_common/sanitizer_stacktrace.h"
 #include "tsan_platform.h"
 #include "tsan_rtl.h"
 #include "tsan_suppressions.h"
@@ -29,12 +30,15 @@ namespace __tsan {
 
 using namespace __sanitizer;  // NOLINT
 
+static ReportStack *SymbolizeStack(const StackTrace& trace);
+
 void TsanCheckFailed(const char *file, int line, const char *cond,
                      u64 v1, u64 v2) {
   ScopedInRtl in_rtl;
   Printf("FATAL: ThreadSanitizer CHECK failed: "
          "%s:%d \"%s\" (0x%zx, 0x%zx)\n",
          file, line, cond, (uptr)v1, (uptr)v2);
+  PrintCurrentStackSlow();
   Die();
 }
 
@@ -121,14 +125,16 @@ static ReportStack *SymbolizeStack(const StackTrace& trace) {
 
 ScopedReport::ScopedReport(ReportType typ) {
   ctx_ = CTX();
-  ctx_->thread_mtx.CheckLocked();
+  ctx_->thread_registry->CheckLocked();
   void *mem = internal_alloc(MBlockReport, sizeof(ReportDesc));
   rep_ = new(mem) ReportDesc;
   rep_->typ = typ;
   ctx_->report_mtx.Lock();
+  CommonSanitizerReportMutex.Lock();
 }
 
 ScopedReport::~ScopedReport() {
+  CommonSanitizerReportMutex.Unlock();
   ctx_->report_mtx.Unlock();
   DestroyAndFree(rep_);
 }
@@ -146,7 +152,8 @@ void ScopedReport::AddMemoryAccess(uptr addr, Shadow s,
   mop->tid = s.tid();
   mop->addr = addr + s.addr0();
   mop->size = s.size();
-  mop->write = s.is_write();
+  mop->write = s.IsWrite();
+  mop->atomic = s.IsAtomic();
   mop->stack = SymbolizeStack(*stack);
   for (uptr i = 0; i < mset->Size(); i++) {
     MutexSet::Desc d = mset->Get(i);
@@ -172,7 +179,7 @@ void ScopedReport::AddMemoryAccess(uptr addr, Shadow s,
 
 void ScopedReport::AddThread(const ThreadContext *tctx) {
   for (uptr i = 0; i < rep_->threads.Size(); i++) {
-    if (rep_->threads[i]->id == tctx->tid)
+    if ((u32)rep_->threads[i]->id == tctx->tid)
       return;
   }
   void *mem = internal_alloc(MBlockReportThread, sizeof(ReportThread));
@@ -182,42 +189,65 @@ void ScopedReport::AddThread(const ThreadContext *tctx) {
   rt->pid = tctx->os_id;
   rt->running = (tctx->status == ThreadStatusRunning);
   rt->name = tctx->name ? internal_strdup(tctx->name) : 0;
-  rt->parent_tid = tctx->creation_tid;
+  rt->parent_tid = tctx->parent_tid;
+  rt->stack = 0;
+#ifdef TSAN_GO
   rt->stack = SymbolizeStack(tctx->creation_stack);
+#else
+  uptr ssz = 0;
+  const uptr *stack = StackDepotGet(tctx->creation_stack_id, &ssz);
+  if (stack) {
+    StackTrace trace;
+    trace.Init(stack, ssz);
+    rt->stack = SymbolizeStack(trace);
+  }
+#endif
 }
 
 #ifndef TSAN_GO
-static ThreadContext *FindThread(int unique_id) {
+static ThreadContext *FindThreadByUidLocked(int unique_id) {
   Context *ctx = CTX();
-  ctx->thread_mtx.CheckLocked();
+  ctx->thread_registry->CheckLocked();
   for (unsigned i = 0; i < kMaxTid; i++) {
-    ThreadContext *tctx = ctx->threads[i];
-    if (tctx && tctx->unique_id == unique_id) {
+    ThreadContext *tctx = static_cast<ThreadContext*>(
+        ctx->thread_registry->GetThreadLocked(i));
+    if (tctx && tctx->unique_id == (u32)unique_id) {
       return tctx;
     }
   }
   return 0;
 }
 
+static ThreadContext *FindThreadByTidLocked(int tid) {
+  Context *ctx = CTX();
+  ctx->thread_registry->CheckLocked();
+  return static_cast<ThreadContext*>(
+      ctx->thread_registry->GetThreadLocked(tid));
+}
+
+static bool IsInStackOrTls(ThreadContextBase *tctx_base, void *arg) {
+  uptr addr = (uptr)arg;
+  ThreadContext *tctx = static_cast<ThreadContext*>(tctx_base);
+  if (tctx->status != ThreadStatusRunning)
+    return false;
+  ThreadState *thr = tctx->thr;
+  CHECK(thr);
+  return ((addr >= thr->stk_addr && addr < thr->stk_addr + thr->stk_size) ||
+          (addr >= thr->tls_addr && addr < thr->tls_addr + thr->tls_size));
+}
+
 ThreadContext *IsThreadStackOrTls(uptr addr, bool *is_stack) {
   Context *ctx = CTX();
-  ctx->thread_mtx.CheckLocked();
-  for (unsigned i = 0; i < kMaxTid; i++) {
-    ThreadContext *tctx = ctx->threads[i];
-    if (tctx == 0 || tctx->status != ThreadStatusRunning)
-      continue;
-    ThreadState *thr = tctx->thr;
-    CHECK(thr);
-    if (addr >= thr->stk_addr && addr < thr->stk_addr + thr->stk_size) {
-      *is_stack = true;
-      return tctx;
-    }
-    if (addr >= thr->tls_addr && addr < thr->tls_addr + thr->tls_size) {
-      *is_stack = false;
-      return tctx;
-    }
-  }
-  return 0;
+  ctx->thread_registry->CheckLocked();
+  ThreadContext *tctx = static_cast<ThreadContext*>(
+      ctx->thread_registry->FindThreadContextLocked(IsInStackOrTls,
+                                                    (void*)addr));
+  if (!tctx)
+    return 0;
+  ThreadState *thr = tctx->thr;
+  CHECK(thr);
+  *is_stack = (addr >= thr->stk_addr && addr < thr->stk_addr + thr->stk_size);
+  return tctx;
 }
 #endif
 
@@ -231,7 +261,16 @@ void ScopedReport::AddMutex(const SyncVar *s) {
   rep_->mutexes.PushBack(rm);
   rm->id = s->uid;
   rm->destroyed = false;
-  rm->stack = SymbolizeStack(s->creation_stack);
+  rm->stack = 0;
+#ifndef TSAN_GO
+  uptr ssz = 0;
+  const uptr *stack = StackDepotGet(s->creation_stack_id, &ssz);
+  if (stack) {
+    StackTrace trace;
+    trace.Init(stack, ssz);
+    rm->stack = SymbolizeStack(trace);
+  }
+#endif
 }
 
 void ScopedReport::AddMutex(u64 id) {
@@ -269,27 +308,28 @@ void ScopedReport::AddLocation(uptr addr, uptr size) {
       trace.Init(stack, ssz);
       loc->stack = SymbolizeStack(trace);
     }
-    ThreadContext *tctx = FindThread(creat_tid);
+    ThreadContext *tctx = FindThreadByUidLocked(creat_tid);
     if (tctx)
       AddThread(tctx);
     return;
   }
-  if (allocator()->PointerIsMine((void*)addr)) {
-    MBlock *b = user_mblock(0, (void*)addr);
-    ThreadContext *tctx = FindThread(b->alloc_tid);
+  MBlock *b = 0;
+  if (allocator()->PointerIsMine((void*)addr)
+      && (b = user_mblock(0, (void*)addr))) {
+    ThreadContext *tctx = FindThreadByTidLocked(b->Tid());
     void *mem = internal_alloc(MBlockReportLoc, sizeof(ReportLocation));
     ReportLocation *loc = new(mem) ReportLocation();
     rep_->locs.PushBack(loc);
     loc->type = ReportLocationHeap;
     loc->addr = (uptr)allocator()->GetBlockBegin((void*)addr);
-    loc->size = b->size;
-    loc->tid = tctx ? tctx->tid : b->alloc_tid;
+    loc->size = b->Size();
+    loc->tid = tctx ? tctx->tid : b->Tid();
     loc->name = 0;
     loc->file = 0;
     loc->line = 0;
     loc->stack = 0;
     uptr ssz = 0;
-    const uptr *stack = StackDepotGet(b->alloc_stack_id, &ssz);
+    const uptr *stack = StackDepotGet(b->StackId(), &ssz);
     if (stack) {
       StackTrace trace;
       trace.Init(stack, ssz);
@@ -328,6 +368,10 @@ void ScopedReport::AddSleep(u32 stack_id) {
 }
 #endif
 
+void ScopedReport::SetCount(int count) {
+  rep_->count = count;
+}
+
 const ReportDesc *ScopedReport::GetReport() const {
   return rep_;
 }
@@ -336,21 +380,17 @@ void RestoreStack(int tid, const u64 epoch, StackTrace *stk, MutexSet *mset) {
   // This function restores stack trace and mutex set for the thread/epoch.
   // It does so by getting stack trace and mutex set at the beginning of
   // trace part, and then replaying the trace till the given epoch.
-  ThreadContext *tctx = CTX()->threads[tid];
+  Context *ctx = CTX();
+  ctx->thread_registry->CheckLocked();
+  ThreadContext *tctx = static_cast<ThreadContext*>(
+      ctx->thread_registry->GetThreadLocked(tid));
   if (tctx == 0)
     return;
-  Trace* trace = 0;
-  if (tctx->status == ThreadStatusRunning) {
-    CHECK(tctx->thr);
-    trace = &tctx->thr->trace;
-  } else if (tctx->status == ThreadStatusFinished
-      || tctx->status == ThreadStatusDead) {
-    if (tctx->dead_info == 0)
-      return;
-    trace = &tctx->dead_info->trace;
-  } else {
+  if (tctx->status != ThreadStatusRunning
+      && tctx->status != ThreadStatusFinished
+      && tctx->status != ThreadStatusDead)
     return;
-  }
+  Trace* trace = ThreadTrace(tctx->tid);
   Lock l(&trace->mtx);
   const int partidx = (epoch / kTracePartSize) % TraceParts();
   TraceHeader* hdr = &trace->headers[partidx];
@@ -460,12 +500,17 @@ static void AddRacyStacks(ThreadState *thr, const StackTrace (&traces)[2],
 
 bool OutputReport(Context *ctx,
                   const ScopedReport &srep,
-                  const ReportStack *suppress_stack) {
+                  const ReportStack *suppress_stack1,
+                  const ReportStack *suppress_stack2) {
+  atomic_store(&ctx->last_symbolize_time_ns, NanoTime(), memory_order_relaxed);
   const ReportDesc *rep = srep.GetReport();
-  const uptr suppress_pc = IsSuppressed(rep->typ, suppress_stack);
+  Suppression *supp = 0;
+  uptr suppress_pc = IsSuppressed(rep->typ, suppress_stack1, &supp);
+  if (suppress_pc == 0)
+    suppress_pc = IsSuppressed(rep->typ, suppress_stack2, &supp);
   if (suppress_pc != 0) {
-    FiredSuppression supp = {srep.GetReport()->typ, suppress_pc};
-    ctx->fired_suppressions.PushBack(supp);
+    FiredSuppression s = {srep.GetReport()->typ, suppress_pc, supp};
+    ctx->fired_suppressions.PushBack(s);
   }
   if (OnReport(rep, suppress_pc != 0))
     return false;
@@ -481,13 +526,24 @@ bool IsFiredSuppression(Context *ctx,
     if (ctx->fired_suppressions[k].type != srep.GetReport()->typ)
       continue;
     for (uptr j = 0; j < trace.Size(); j++) {
-      if (trace.Get(j) == ctx->fired_suppressions[k].pc)
+      FiredSuppression *s = &ctx->fired_suppressions[k];
+      if (trace.Get(j) == s->pc) {
+        if (s->supp)
+          s->supp->hit_count++;
         return true;
+      }
     }
   }
   return false;
 }
 
+bool FrameIsInternal(const ReportStack *frame) {
+  return frame != 0 && frame->file != 0
+      && (internal_strstr(frame->file, "tsan_interceptors.cc") ||
+          internal_strstr(frame->file, "sanitizer_common_interceptors.inc") ||
+          internal_strstr(frame->file, "tsan_interface_"));
+}
+
 // On programs that use Java we see weird reports like:
 // WARNING: ThreadSanitizer: data race (pid=22512)
 //   Read of size 8 at 0x7d2b00084318 by thread 100:
@@ -497,30 +553,42 @@ bool IsFiredSuppression(Context *ctx,
 //     #0 strncpy tsan_interceptors.cc:501 (foo+0x00000d8e0919)
 //     #1 <null> <null>:0 (0x7f7ad9b42707)
 static bool IsJavaNonsense(const ReportDesc *rep) {
+#ifndef TSAN_GO
   for (uptr i = 0; i < rep->mops.Size(); i++) {
     ReportMop *mop = rep->mops[i];
     ReportStack *frame = mop->stack;
-    if (frame != 0 && frame->func != 0
-        && (internal_strcmp(frame->func, "memset") == 0
-        || internal_strcmp(frame->func, "memcpy") == 0
-        || internal_strcmp(frame->func, "memmove") == 0
-        || internal_strcmp(frame->func, "strcmp") == 0
-        || internal_strcmp(frame->func, "strncpy") == 0
-        || internal_strcmp(frame->func, "strlen") == 0
-        || internal_strcmp(frame->func, "free") == 0
-        || internal_strcmp(frame->func, "pthread_mutex_lock") == 0)) {
+    if (frame == 0
+        || (frame->func == 0 && frame->file == 0 && frame->line == 0
+          && frame->module == 0)) {
+      return true;
+    }
+    if (FrameIsInternal(frame)) {
       frame = frame->next;
       if (frame == 0
           || (frame->func == 0 && frame->file == 0 && frame->line == 0
-            && frame->module == 0)) {
+          && frame->module == 0)) {
         if (frame) {
-          FiredSuppression supp = {rep->typ, frame->pc};
+          FiredSuppression supp = {rep->typ, frame->pc, 0};
           CTX()->fired_suppressions.PushBack(supp);
         }
         return true;
       }
     }
   }
+#endif
+  return false;
+}
+
+static bool RaceBetweenAtomicAndFree(ThreadState *thr) {
+  Shadow s0(thr->racy_state[0]);
+  Shadow s1(thr->racy_state[1]);
+  CHECK(!(s0.IsAtomic() && s1.IsAtomic()));
+  if (!s0.IsAtomic() && !s1.IsAtomic())
+    return true;
+  if (s0.IsAtomic() && s1.IsFreed())
+    return true;
+  if (s1.IsAtomic() && thr->is_freeing)
+    return true;
   return false;
 }
 
@@ -529,9 +597,8 @@ void ReportRace(ThreadState *thr) {
     return;
   ScopedInRtl in_rtl;
 
-  if (thr->in_signal_handler)
-    Printf("ThreadSanitizer: printing report from signal handler."
-           " Can crash or hang.\n");
+  if (!flags()->report_atomic_races && !RaceBetweenAtomicAndFree(thr))
+    return;
 
   bool freed = false;
   {
@@ -555,9 +622,14 @@ void ReportRace(ThreadState *thr) {
   }
 
   Context *ctx = CTX();
-  Lock l0(&ctx->thread_mtx);
-
-  ScopedReport rep(freed ? ReportTypeUseAfterFree : ReportTypeRace);
+  ThreadRegistryLock l0(ctx->thread_registry);
+
+  ReportType typ = ReportTypeRace;
+  if (thr->is_vptr_access)
+    typ = ReportTypeVptrRace;
+  else if (freed)
+    typ = ReportTypeUseAfterFree;
+  ScopedReport rep(typ);
   const uptr kMop = 2;
   StackTrace traces[kMop];
   const uptr toppc = TraceTopPC(thr);
@@ -583,7 +655,8 @@ void ReportRace(ThreadState *thr) {
 
   for (uptr i = 0; i < kMop; i++) {
     FastState s(thr->racy_state[i]);
-    ThreadContext *tctx = ctx->threads[s.tid()];
+    ThreadContext *tctx = static_cast<ThreadContext*>(
+        ctx->thread_registry->GetThreadLocked(s.tid()));
     if (s.epoch() < tctx->epoch0 || s.epoch() > tctx->epoch1)
       continue;
     rep.AddThread(tctx);
@@ -599,7 +672,8 @@ void ReportRace(ThreadState *thr) {
   }
 #endif
 
-  if (!OutputReport(ctx, rep, rep.GetReport()->mops[0]->stack))
+  if (!OutputReport(ctx, rep, rep.GetReport()->mops[0]->stack,
+                              rep.GetReport()->mops[1]->stack))
     return;
 
   AddRacyStacks(thr, traces, addr_min, addr_max);
@@ -611,4 +685,21 @@ void PrintCurrentStack(ThreadState *thr, uptr pc) {
   PrintStack(SymbolizeStack(trace));
 }
 
+void PrintCurrentStackSlow() {
+#ifndef TSAN_GO
+  __sanitizer::StackTrace *ptrace = new(internal_alloc(MBlockStackTrace,
+      sizeof(__sanitizer::StackTrace))) __sanitizer::StackTrace;
+  ptrace->SlowUnwindStack(__sanitizer::StackTrace::GetCurrentPc(),
+      kStackTraceMax);
+  for (uptr i = 0; i < ptrace->size / 2; i++) {
+    uptr tmp = ptrace->trace[i];
+    ptrace->trace[i] = ptrace->trace[ptrace->size - i - 1];
+    ptrace->trace[ptrace->size - i - 1] = tmp;
+  }
+  StackTrace trace;
+  trace.Init(ptrace->trace, ptrace->size);
+  PrintStack(SymbolizeStack(trace));
+#endif
+}
+
 }  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_rtl_thread.cc b/lib/tsan/rtl/tsan_rtl_thread.cc
index 3597759278345..ee13fa18db3f9 100644
--- a/lib/tsan/rtl/tsan_rtl_thread.cc
+++ b/lib/tsan/rtl/tsan_rtl_thread.cc
@@ -20,144 +20,193 @@
 
 namespace __tsan {
 
+// ThreadContext implementation.
+
+ThreadContext::ThreadContext(int tid)
+  : ThreadContextBase(tid)
+  , thr()
+  , sync()
+  , epoch0()
+  , epoch1() {
+}
+
 #ifndef TSAN_GO
-const int kThreadQuarantineSize = 16;
-#else
-const int kThreadQuarantineSize = 64;
+ThreadContext::~ThreadContext() {
+}
 #endif
 
-static void MaybeReportThreadLeak(ThreadContext *tctx) {
-  if (tctx->detached)
+void ThreadContext::OnDead() {
+  sync.Reset();
+}
+
+void ThreadContext::OnJoined(void *arg) {
+  ThreadState *caller_thr = static_cast<ThreadState *>(arg);
+  caller_thr->clock.acquire(&sync);
+  StatInc(caller_thr, StatSyncAcquire);
+  sync.Reset();
+}
+
+struct OnCreatedArgs {
+  ThreadState *thr;
+  uptr pc;
+};
+
+void ThreadContext::OnCreated(void *arg) {
+  thr = 0;
+  if (tid == 0)
     return;
-  if (tctx->status != ThreadStatusCreated
-      && tctx->status != ThreadStatusRunning
-      && tctx->status != ThreadStatusFinished)
+  OnCreatedArgs *args = static_cast<OnCreatedArgs *>(arg);
+  args->thr->fast_state.IncrementEpoch();
+  // Can't increment epoch w/o writing to the trace as well.
+  TraceAddEvent(args->thr, args->thr->fast_state, EventTypeMop, 0);
+  args->thr->clock.set(args->thr->tid, args->thr->fast_state.epoch());
+  args->thr->fast_synch_epoch = args->thr->fast_state.epoch();
+  args->thr->clock.release(&sync);
+  StatInc(args->thr, StatSyncRelease);
+#ifdef TSAN_GO
+  creation_stack.ObtainCurrent(args->thr, args->pc);
+#else
+  creation_stack_id = CurrentStackId(args->thr, args->pc);
+#endif
+  if (reuse_count == 0)
+    StatInc(args->thr, StatThreadMaxTid);
+}
+
+void ThreadContext::OnReset() {
+  sync.Reset();
+  FlushUnneededShadowMemory(GetThreadTrace(tid), TraceSize() * sizeof(Event));
+  //!!! FlushUnneededShadowMemory(GetThreadTraceHeader(tid), sizeof(Trace));
+}
+
+struct OnStartedArgs {
+  ThreadState *thr;
+  uptr stk_addr;
+  uptr stk_size;
+  uptr tls_addr;
+  uptr tls_size;
+};
+
+void ThreadContext::OnStarted(void *arg) {
+  OnStartedArgs *args = static_cast<OnStartedArgs*>(arg);
+  thr = args->thr;
+  // RoundUp so that one trace part does not contain events
+  // from different threads.
+  epoch0 = RoundUp(epoch1 + 1, kTracePartSize);
+  epoch1 = (u64)-1;
+  new(thr) ThreadState(CTX(), tid, unique_id,
+      epoch0, args->stk_addr, args->stk_size, args->tls_addr, args->tls_size);
+#ifdef TSAN_GO
+  // Setup dynamic shadow stack.
+  const int kInitStackSize = 8;
+  args->thr->shadow_stack = (uptr*)internal_alloc(MBlockShadowStack,
+      kInitStackSize * sizeof(uptr));
+  args->thr->shadow_stack_pos = thr->shadow_stack;
+  args->thr->shadow_stack_end = thr->shadow_stack + kInitStackSize;
+#endif
+#ifndef TSAN_GO
+  AllocatorThreadStart(args->thr);
+#endif
+  thr = args->thr;
+  thr->fast_synch_epoch = epoch0;
+  thr->clock.set(tid, epoch0);
+  thr->clock.acquire(&sync);
+  thr->fast_state.SetHistorySize(flags()->history_size);
+  const uptr trace = (epoch0 / kTracePartSize) % TraceParts();
+  Trace *thr_trace = ThreadTrace(thr->tid);
+  thr_trace->headers[trace].epoch0 = epoch0;
+  StatInc(thr, StatSyncAcquire);
+  sync.Reset();
+  DPrintf("#%d: ThreadStart epoch=%zu stk_addr=%zx stk_size=%zx "
+          "tls_addr=%zx tls_size=%zx\n",
+          tid, (uptr)epoch0, args->stk_addr, args->stk_size,
+          args->tls_addr, args->tls_size);
+  thr->is_alive = true;
+}
+
+void ThreadContext::OnFinished() {
+  if (!detached) {
+    thr->fast_state.IncrementEpoch();
+    // Can't increment epoch w/o writing to the trace as well.
+    TraceAddEvent(thr, thr->fast_state, EventTypeMop, 0);
+    thr->clock.set(thr->tid, thr->fast_state.epoch());
+    thr->fast_synch_epoch = thr->fast_state.epoch();
+    thr->clock.release(&sync);
+    StatInc(thr, StatSyncRelease);
+  }
+  epoch1 = thr->fast_state.epoch();
+
+#ifndef TSAN_GO
+  AllocatorThreadFinish(thr);
+#endif
+  thr->~ThreadState();
+  StatAggregate(CTX()->stat, thr->stat);
+  thr = 0;
+}
+
+#ifndef TSAN_GO
+struct ThreadLeak {
+  ThreadContext *tctx;
+  int count;
+};
+
+static void MaybeReportThreadLeak(ThreadContextBase *tctx_base, void *arg) {
+  Vector<ThreadLeak> &leaks = *(Vector<ThreadLeak>*)arg;
+  ThreadContext *tctx = static_cast<ThreadContext*>(tctx_base);
+  if (tctx->detached || tctx->status != ThreadStatusFinished)
     return;
-  ScopedReport rep(ReportTypeThreadLeak);
-  rep.AddThread(tctx);
-  OutputReport(CTX(), rep);
+  for (uptr i = 0; i < leaks.Size(); i++) {
+    if (leaks[i].tctx->creation_stack_id == tctx->creation_stack_id) {
+      leaks[i].count++;
+      return;
+    }
+  }
+  ThreadLeak leak = {tctx, 1};
+  leaks.PushBack(leak);
+}
+#endif
+
+static void ThreadCheckIgnore(ThreadState *thr) {
+  if (thr->ignore_reads_and_writes) {
+    Printf("ThreadSanitizer: thread T%d finished with ignores enabled.\n",
+           thr->tid);
+  }
 }
 
 void ThreadFinalize(ThreadState *thr) {
   CHECK_GT(thr->in_rtl, 0);
+  ThreadCheckIgnore(thr);
+#ifndef TSAN_GO
   if (!flags()->report_thread_leaks)
     return;
-  Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  for (unsigned i = 0; i < kMaxTid; i++) {
-    ThreadContext *tctx = ctx->threads[i];
-    if (tctx == 0)
-      continue;
-    MaybeReportThreadLeak(tctx);
+  ThreadRegistryLock l(CTX()->thread_registry);
+  Vector<ThreadLeak> leaks(MBlockScopedBuf);
+  CTX()->thread_registry->RunCallbackForEachThreadLocked(
+      MaybeReportThreadLeak, &leaks);
+  for (uptr i = 0; i < leaks.Size(); i++) {
+    ScopedReport rep(ReportTypeThreadLeak);
+    rep.AddThread(leaks[i].tctx);
+    rep.SetCount(leaks[i].count);
+    OutputReport(CTX(), rep);
   }
+#endif
 }
 
 int ThreadCount(ThreadState *thr) {
   CHECK_GT(thr->in_rtl, 0);
   Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  int cnt = 0;
-  for (unsigned i = 0; i < kMaxTid; i++) {
-    ThreadContext *tctx = ctx->threads[i];
-    if (tctx == 0)
-      continue;
-    if (tctx->status != ThreadStatusCreated
-        && tctx->status != ThreadStatusRunning)
-      continue;
-    cnt++;
-  }
-  return cnt;
-}
-
-static void ThreadDead(ThreadState *thr, ThreadContext *tctx) {
-  Context *ctx = CTX();
-  CHECK_GT(thr->in_rtl, 0);
-  CHECK(tctx->status == ThreadStatusRunning
-      || tctx->status == ThreadStatusFinished);
-  DPrintf("#%d: ThreadDead uid=%zu\n", thr->tid, tctx->user_id);
-  tctx->status = ThreadStatusDead;
-  tctx->user_id = 0;
-  tctx->sync.Reset();
-
-  // Put to dead list.
-  tctx->dead_next = 0;
-  if (ctx->dead_list_size == 0)
-    ctx->dead_list_head = tctx;
-  else
-    ctx->dead_list_tail->dead_next = tctx;
-  ctx->dead_list_tail = tctx;
-  ctx->dead_list_size++;
+  uptr result;
+  ctx->thread_registry->GetNumberOfThreads(0, 0, &result);
+  return (int)result;
 }
 
 int ThreadCreate(ThreadState *thr, uptr pc, uptr uid, bool detached) {
   CHECK_GT(thr->in_rtl, 0);
-  Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
   StatInc(thr, StatThreadCreate);
-  int tid = -1;
-  ThreadContext *tctx = 0;
-  if (ctx->dead_list_size > kThreadQuarantineSize
-      || ctx->thread_seq >= kMaxTid) {
-    // Reusing old thread descriptor and tid.
-    if (ctx->dead_list_size == 0) {
-      Printf("ThreadSanitizer: %d thread limit exceeded. Dying.\n",
-                 kMaxTid);
-      Die();
-    }
-    StatInc(thr, StatThreadReuse);
-    tctx = ctx->dead_list_head;
-    ctx->dead_list_head = tctx->dead_next;
-    ctx->dead_list_size--;
-    if (ctx->dead_list_size == 0) {
-      CHECK_EQ(tctx->dead_next, 0);
-      ctx->dead_list_head = 0;
-    }
-    CHECK_EQ(tctx->status, ThreadStatusDead);
-    tctx->status = ThreadStatusInvalid;
-    tctx->reuse_count++;
-    tctx->sync.Reset();
-    tid = tctx->tid;
-    DestroyAndFree(tctx->dead_info);
-    if (tctx->name) {
-      internal_free(tctx->name);
-      tctx->name = 0;
-    }
-  } else {
-    // Allocating new thread descriptor and tid.
-    StatInc(thr, StatThreadMaxTid);
-    tid = ctx->thread_seq++;
-    void *mem = internal_alloc(MBlockThreadContex, sizeof(ThreadContext));
-    tctx = new(mem) ThreadContext(tid);
-    ctx->threads[tid] = tctx;
-    MapThreadTrace(GetThreadTrace(tid), TraceSize() * sizeof(Event));
-  }
-  CHECK_NE(tctx, 0);
-  CHECK_GE(tid, 0);
-  CHECK_LT(tid, kMaxTid);
+  Context *ctx = CTX();
+  OnCreatedArgs args = { thr, pc };
+  int tid = ctx->thread_registry->CreateThread(uid, detached, thr->tid, &args);
   DPrintf("#%d: ThreadCreate tid=%d uid=%zu\n", thr->tid, tid, uid);
-  CHECK_EQ(tctx->status, ThreadStatusInvalid);
-  ctx->alive_threads++;
-  if (ctx->max_alive_threads < ctx->alive_threads) {
-    ctx->max_alive_threads++;
-    CHECK_EQ(ctx->max_alive_threads, ctx->alive_threads);
-    StatInc(thr, StatThreadMaxAlive);
-  }
-  tctx->status = ThreadStatusCreated;
-  tctx->thr = 0;
-  tctx->user_id = uid;
-  tctx->unique_id = ctx->unique_thread_seq++;
-  tctx->detached = detached;
-  if (tid) {
-    thr->fast_state.IncrementEpoch();
-    // Can't increment epoch w/o writing to the trace as well.
-    TraceAddEvent(thr, thr->fast_state, EventTypeMop, 0);
-    thr->clock.set(thr->tid, thr->fast_state.epoch());
-    thr->fast_synch_epoch = thr->fast_state.epoch();
-    thr->clock.release(&tctx->sync);
-    StatInc(thr, StatSyncRelease);
-    tctx->creation_stack.ObtainCurrent(thr, pc);
-    tctx->creation_tid = thr->tid;
-  }
+  StatSet(thr, StatThreadMaxAlive, ctx->thread_registry->GetMaxAliveThreads());
   return tid;
 }
 
@@ -170,9 +219,8 @@ void ThreadStart(ThreadState *thr, int tid, uptr os_id) {
   GetThreadStackAndTls(tid == 0, &stk_addr, &stk_size, &tls_addr, &tls_size);
 
   if (tid) {
-    if (stk_addr && stk_size) {
-      MemoryResetRange(thr, /*pc=*/ 1, stk_addr, stk_size);
-    }
+    if (stk_addr && stk_size)
+      MemoryRangeImitateWrite(thr, /*pc=*/ 1, stk_addr, stk_size);
 
     if (tls_addr && tls_size) {
       // Check that the thr object is in tls;
@@ -183,113 +231,42 @@ void ThreadStart(ThreadState *thr, int tid, uptr os_id) {
       CHECK_GE(thr_end, tls_addr);
       CHECK_LE(thr_end, tls_addr + tls_size);
       // Since the thr object is huge, skip it.
-      MemoryResetRange(thr, /*pc=*/ 2, tls_addr, thr_beg - tls_addr);
-      MemoryResetRange(thr, /*pc=*/ 2, thr_end, tls_addr + tls_size - thr_end);
+      MemoryRangeImitateWrite(thr, /*pc=*/ 2, tls_addr, thr_beg - tls_addr);
+      MemoryRangeImitateWrite(thr, /*pc=*/ 2,
+          thr_end, tls_addr + tls_size - thr_end);
     }
   }
 
-  Lock l(&CTX()->thread_mtx);
-  ThreadContext *tctx = CTX()->threads[tid];
-  CHECK_NE(tctx, 0);
-  CHECK_EQ(tctx->status, ThreadStatusCreated);
-  tctx->status = ThreadStatusRunning;
-  tctx->os_id = os_id;
-  // RoundUp so that one trace part does not contain events
-  // from different threads.
-  tctx->epoch0 = RoundUp(tctx->epoch1 + 1, kTracePartSize);
-  tctx->epoch1 = (u64)-1;
-  new(thr) ThreadState(CTX(), tid, tctx->unique_id,
-      tctx->epoch0, stk_addr, stk_size,
-      tls_addr, tls_size);
-#ifdef TSAN_GO
-  // Setup dynamic shadow stack.
-  const int kInitStackSize = 8;
-  thr->shadow_stack = (uptr*)internal_alloc(MBlockShadowStack,
-      kInitStackSize * sizeof(uptr));
-  thr->shadow_stack_pos = thr->shadow_stack;
-  thr->shadow_stack_end = thr->shadow_stack + kInitStackSize;
-#endif
-  tctx->thr = thr;
-  thr->fast_synch_epoch = tctx->epoch0;
-  thr->clock.set(tid, tctx->epoch0);
-  thr->clock.acquire(&tctx->sync);
-  thr->fast_state.SetHistorySize(flags()->history_size);
-  const uptr trace = (tctx->epoch0 / kTracePartSize) % TraceParts();
-  thr->trace.headers[trace].epoch0 = tctx->epoch0;
-  StatInc(thr, StatSyncAcquire);
-  DPrintf("#%d: ThreadStart epoch=%zu stk_addr=%zx stk_size=%zx "
-          "tls_addr=%zx tls_size=%zx\n",
-          tid, (uptr)tctx->epoch0, stk_addr, stk_size, tls_addr, tls_size);
-  thr->is_alive = true;
+  OnStartedArgs args = { thr, stk_addr, stk_size, tls_addr, tls_size };
+  CTX()->thread_registry->StartThread(tid, os_id, &args);
 }
 
 void ThreadFinish(ThreadState *thr) {
   CHECK_GT(thr->in_rtl, 0);
+  ThreadCheckIgnore(thr);
   StatInc(thr, StatThreadFinish);
-  // FIXME: Treat it as write.
   if (thr->stk_addr && thr->stk_size)
-    MemoryResetRange(thr, /*pc=*/ 3, thr->stk_addr, thr->stk_size);
-  if (thr->tls_addr && thr->tls_size) {
-    const uptr thr_beg = (uptr)thr;
-    const uptr thr_end = (uptr)thr + sizeof(*thr);
-    // Since the thr object is huge, skip it.
-    MemoryResetRange(thr, /*pc=*/ 4, thr->tls_addr, thr_beg - thr->tls_addr);
-    MemoryResetRange(thr, /*pc=*/ 5,
-        thr_end, thr->tls_addr + thr->tls_size - thr_end);
-  }
+    DontNeedShadowFor(thr->stk_addr, thr->stk_size);
+  if (thr->tls_addr && thr->tls_size)
+    DontNeedShadowFor(thr->tls_addr, thr->tls_size);
   thr->is_alive = false;
   Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  ThreadContext *tctx = ctx->threads[thr->tid];
-  CHECK_NE(tctx, 0);
-  CHECK_EQ(tctx->status, ThreadStatusRunning);
-  CHECK_GT(ctx->alive_threads, 0);
-  ctx->alive_threads--;
-  if (tctx->detached) {
-    ThreadDead(thr, tctx);
-  } else {
-    thr->fast_state.IncrementEpoch();
-    // Can't increment epoch w/o writing to the trace as well.
-    TraceAddEvent(thr, thr->fast_state, EventTypeMop, 0);
-    thr->clock.set(thr->tid, thr->fast_state.epoch());
-    thr->fast_synch_epoch = thr->fast_state.epoch();
-    thr->clock.release(&tctx->sync);
-    StatInc(thr, StatSyncRelease);
-    tctx->status = ThreadStatusFinished;
-  }
+  ctx->thread_registry->FinishThread(thr->tid);
+}
 
-  // Save from info about the thread.
-  tctx->dead_info = new(internal_alloc(MBlockDeadInfo, sizeof(ThreadDeadInfo)))
-      ThreadDeadInfo();
-  for (uptr i = 0; i < TraceParts(); i++) {
-    tctx->dead_info->trace.headers[i].epoch0 = thr->trace.headers[i].epoch0;
-    tctx->dead_info->trace.headers[i].stack0.CopyFrom(
-        thr->trace.headers[i].stack0);
+static bool FindThreadByUid(ThreadContextBase *tctx, void *arg) {
+  uptr uid = (uptr)arg;
+  if (tctx->user_id == uid && tctx->status != ThreadStatusInvalid) {
+    tctx->user_id = 0;
+    return true;
   }
-  tctx->epoch1 = thr->fast_state.epoch();
-
-#ifndef TSAN_GO
-  AlloctorThreadFinish(thr);
-#endif
-  thr->~ThreadState();
-  StatAggregate(ctx->stat, thr->stat);
-  tctx->thr = 0;
+  return false;
 }
 
 int ThreadTid(ThreadState *thr, uptr pc, uptr uid) {
   CHECK_GT(thr->in_rtl, 0);
   Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  int res = -1;
-  for (unsigned tid = 0; tid < kMaxTid; tid++) {
-    ThreadContext *tctx = ctx->threads[tid];
-    if (tctx != 0 && tctx->user_id == uid
-        && tctx->status != ThreadStatusInvalid) {
-      tctx->user_id = 0;
-      res = tid;
-      break;
-    }
-  }
+  int res = ctx->thread_registry->FindThread(FindThreadByUid, (void*)uid);
   DPrintf("#%d: ThreadTid uid=%zu tid=%d\n", thr->tid, uid, res);
   return res;
 }
@@ -300,18 +277,7 @@ void ThreadJoin(ThreadState *thr, uptr pc, int tid) {
   CHECK_LT(tid, kMaxTid);
   DPrintf("#%d: ThreadJoin tid=%d\n", thr->tid, tid);
   Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  ThreadContext *tctx = ctx->threads[tid];
-  if (tctx->status == ThreadStatusInvalid) {
-    Printf("ThreadSanitizer: join of non-existent thread\n");
-    return;
-  }
-  // FIXME(dvyukov): print message and continue (it's user error).
-  CHECK_EQ(tctx->detached, false);
-  CHECK_EQ(tctx->status, ThreadStatusFinished);
-  thr->clock.acquire(&tctx->sync);
-  StatInc(thr, StatSyncAcquire);
-  ThreadDead(thr, tctx);
+  ctx->thread_registry->JoinThread(tid, thr);
 }
 
 void ThreadDetach(ThreadState *thr, uptr pc, int tid) {
@@ -319,31 +285,12 @@ void ThreadDetach(ThreadState *thr, uptr pc, int tid) {
   CHECK_GT(tid, 0);
   CHECK_LT(tid, kMaxTid);
   Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  ThreadContext *tctx = ctx->threads[tid];
-  if (tctx->status == ThreadStatusInvalid) {
-    Printf("ThreadSanitizer: detach of non-existent thread\n");
-    return;
-  }
-  if (tctx->status == ThreadStatusFinished) {
-    ThreadDead(thr, tctx);
-  } else {
-    tctx->detached = true;
-  }
+  ctx->thread_registry->DetachThread(tid);
 }
 
 void ThreadSetName(ThreadState *thr, const char *name) {
-  Context *ctx = CTX();
-  Lock l(&ctx->thread_mtx);
-  ThreadContext *tctx = ctx->threads[thr->tid];
-  CHECK_NE(tctx, 0);
-  CHECK_EQ(tctx->status, ThreadStatusRunning);
-  if (tctx->name) {
-    internal_free(tctx->name);
-    tctx->name = 0;
-  }
-  if (name)
-    tctx->name = internal_strdup(name);
+  CHECK_GT(thr->in_rtl, 0);
+  CTX()->thread_registry->SetThreadName(thr->tid, name);
 }
 
 void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
@@ -378,6 +325,13 @@ void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
 
   StatInc(thr, StatMopRange);
 
+  if (*shadow_mem == kShadowRodata) {
+    // Access to .rodata section, no races here.
+    // Measurements show that it can be 10-20% of all memory accesses.
+    StatInc(thr, StatMopRangeRodata);
+    return;
+  }
+
   FastState fast_state = thr->fast_state;
   if (fast_state.GetIgnoreBit())
     return;
@@ -394,7 +348,7 @@ void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
     Shadow cur(fast_state);
     cur.SetWrite(is_write);
     cur.SetAddr0AndSizeLog(addr & (kShadowCell - 1), kAccessSizeLog);
-    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write,
+    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write, false,
         shadow_mem, cur);
   }
   if (unaligned)
@@ -405,7 +359,7 @@ void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
     Shadow cur(fast_state);
     cur.SetWrite(is_write);
     cur.SetAddr0AndSizeLog(0, kAccessSizeLog);
-    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write,
+    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write, false,
         shadow_mem, cur);
     shadow_mem += kShadowCnt;
   }
@@ -415,24 +369,30 @@ void MemoryAccessRange(ThreadState *thr, uptr pc, uptr addr,
     Shadow cur(fast_state);
     cur.SetWrite(is_write);
     cur.SetAddr0AndSizeLog(addr & (kShadowCell - 1), kAccessSizeLog);
-    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write,
+    MemoryAccessImpl(thr, addr, kAccessSizeLog, is_write, false,
         shadow_mem, cur);
   }
 }
 
-void MemoryRead1Byte(ThreadState *thr, uptr pc, uptr addr) {
-  MemoryAccess(thr, pc, addr, 0, 0);
-}
-
-void MemoryWrite1Byte(ThreadState *thr, uptr pc, uptr addr) {
-  MemoryAccess(thr, pc, addr, 0, 1);
-}
-
-void MemoryRead8Byte(ThreadState *thr, uptr pc, uptr addr) {
-  MemoryAccess(thr, pc, addr, 3, 0);
-}
+void MemoryAccessRangeStep(ThreadState *thr, uptr pc, uptr addr,
+    uptr size, uptr step, bool is_write) {
+  if (size == 0)
+    return;
+  FastState fast_state = thr->fast_state;
+  if (fast_state.GetIgnoreBit())
+    return;
+  StatInc(thr, StatMopRange);
+  fast_state.IncrementEpoch();
+  thr->fast_state = fast_state;
+  TraceAddEvent(thr, fast_state, EventTypeMop, pc);
 
-void MemoryWrite8Byte(ThreadState *thr, uptr pc, uptr addr) {
-  MemoryAccess(thr, pc, addr, 3, 1);
+  for (uptr addr_end = addr + size; addr < addr_end; addr += step) {
+    u64 *shadow_mem = (u64*)MemToShadow(addr);
+    Shadow cur(fast_state);
+    cur.SetWrite(is_write);
+    cur.SetAddr0AndSizeLog(addr & (kShadowCell - 1), kSizeLog1);
+    MemoryAccessImpl(thr, addr, kSizeLog1, is_write, false,
+        shadow_mem, cur);
+  }
 }
 }  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_stat.cc b/lib/tsan/rtl/tsan_stat.cc
index 82f1d6b5620fd..9676e0872e088 100644
--- a/lib/tsan/rtl/tsan_stat.cc
+++ b/lib/tsan/rtl/tsan_stat.cc
@@ -38,6 +38,8 @@ void StatOutput(u64 *stat) {
   name[StatMop8]                         = "            size 8                ";
   name[StatMopSame]                      = "  Including same                  ";
   name[StatMopRange]                     = "  Including range                 ";
+  name[StatMopRodata]                    = "  Including .rodata               ";
+  name[StatMopRangeRodata]               = "  Including .rodata range         ";
   name[StatShadowProcessed]              = "Shadow processed                  ";
   name[StatShadowZero]                   = "  Including empty                 ";
   name[StatShadowNonZero]                = "  Including non empty             ";
@@ -105,6 +107,7 @@ void StatOutput(u64 *stat) {
   name[StatInt_realloc]                  = "  realloc                         ";
   name[StatInt_free]                     = "  free                            ";
   name[StatInt_cfree]                    = "  cfree                           ";
+  name[StatInt_malloc_usable_size]       = "  malloc_usable_size              ";
   name[StatInt_mmap]                     = "  mmap                            ";
   name[StatInt_mmap64]                   = "  mmap64                          ";
   name[StatInt_munmap]                   = "  munmap                          ";
@@ -135,6 +138,8 @@ void StatOutput(u64 *stat) {
   name[StatInt_strcpy]                   = "  strcpy                          ";
   name[StatInt_strncpy]                  = "  strncpy                         ";
   name[StatInt_strstr]                   = "  strstr                          ";
+  name[StatInt_strcasecmp]               = "  strcasecmp                      ";
+  name[StatInt_strncasecmp]              = "  strncasecmp                     ";
   name[StatInt_atexit]                   = "  atexit                          ";
   name[StatInt___cxa_guard_acquire]      = "  __cxa_guard_acquire             ";
   name[StatInt___cxa_guard_release]      = "  __cxa_guard_release             ";
@@ -174,6 +179,7 @@ void StatOutput(u64 *stat) {
   name[StatInt_pthread_barrier_destroy]  = "  pthread_barrier_destroy         ";
   name[StatInt_pthread_barrier_wait]     = "  pthread_barrier_wait            ";
   name[StatInt_pthread_once]             = "  pthread_once                    ";
+  name[StatInt_pthread_getschedparam]    = "  pthread_getschedparam           ";
   name[StatInt_sem_init]                 = "  sem_init                        ";
   name[StatInt_sem_destroy]              = "  sem_destroy                     ";
   name[StatInt_sem_wait]                 = "  sem_wait                        ";
@@ -181,6 +187,18 @@ void StatOutput(u64 *stat) {
   name[StatInt_sem_timedwait]            = "  sem_timedwait                   ";
   name[StatInt_sem_post]                 = "  sem_post                        ";
   name[StatInt_sem_getvalue]             = "  sem_getvalue                    ";
+  name[StatInt_stat]                     = "  stat                            ";
+  name[StatInt___xstat]                  = "  __xstat                         ";
+  name[StatInt_stat64]                   = "  stat64                          ";
+  name[StatInt___xstat64]                = "  __xstat64                       ";
+  name[StatInt_lstat]                    = "  lstat                           ";
+  name[StatInt___lxstat]                 = "  __lxstat                        ";
+  name[StatInt_lstat64]                  = "  lstat64                         ";
+  name[StatInt___lxstat64]               = "  __lxstat64                      ";
+  name[StatInt_fstat]                    = "  fstat                           ";
+  name[StatInt___fxstat]                 = "  __fxstat                        ";
+  name[StatInt_fstat64]                  = "  fstat64                         ";
+  name[StatInt___fxstat64]               = "  __fxstat64                      ";
   name[StatInt_open]                     = "  open                            ";
   name[StatInt_open64]                   = "  open64                          ";
   name[StatInt_creat]                    = "  creat                           ";
@@ -195,12 +213,15 @@ void StatOutput(u64 *stat) {
   name[StatInt_socket]                   = "  socket                          ";
   name[StatInt_socketpair]               = "  socketpair                      ";
   name[StatInt_connect]                  = "  connect                         ";
+  name[StatInt_bind]                     = "  bind                            ";
+  name[StatInt_listen]                   = "  listen                          ";
   name[StatInt_accept]                   = "  accept                          ";
   name[StatInt_accept4]                  = "  accept4                         ";
   name[StatInt_epoll_create]             = "  epoll_create                    ";
   name[StatInt_epoll_create1]            = "  epoll_create1                   ";
   name[StatInt_close]                    = "  close                           ";
   name[StatInt___close]                  = "  __close                         ";
+  name[StatInt___res_iclose]             = "  __res_iclose                    ";
   name[StatInt_pipe]                     = "  pipe                            ";
   name[StatInt_pipe2]                    = "  pipe2                           ";
   name[StatInt_read]                     = "  read                            ";
@@ -224,6 +245,8 @@ void StatOutput(u64 *stat) {
   name[StatInt_fclose]                   = "  fclose                          ";
   name[StatInt_fread]                    = "  fread                           ";
   name[StatInt_fwrite]                   = "  fwrite                          ";
+  name[StatInt_fflush]                   = "  fflush                          ";
+  name[StatInt_abort]                    = "  abort                           ";
   name[StatInt_puts]                     = "  puts                            ";
   name[StatInt_rmdir]                    = "  rmdir                           ";
   name[StatInt_opendir]                  = "  opendir                         ";
@@ -231,6 +254,10 @@ void StatOutput(u64 *stat) {
   name[StatInt_epoll_wait]               = "  epoll_wait                      ";
   name[StatInt_poll]                     = "  poll                            ";
   name[StatInt_sigaction]                = "  sigaction                       ";
+  name[StatInt_signal]                   = "  signal                          ";
+  name[StatInt_raise]                    = "  raise                           ";
+  name[StatInt_kill]                     = "  kill                            ";
+  name[StatInt_pthread_kill]             = "  pthread_kill                    ";
   name[StatInt_sleep]                    = "  sleep                           ";
   name[StatInt_usleep]                   = "  usleep                          ";
   name[StatInt_nanosleep]                = "  nanosleep                       ";
@@ -242,6 +269,59 @@ void StatOutput(u64 *stat) {
   name[StatInt_scanf]                    = "  scanf                           ";
   name[StatInt_sscanf]                   = "  sscanf                          ";
   name[StatInt_fscanf]                   = "  fscanf                          ";
+  name[StatInt___isoc99_vscanf]          = "  vscanf                          ";
+  name[StatInt___isoc99_vsscanf]         = "  vsscanf                         ";
+  name[StatInt___isoc99_vfscanf]         = "  vfscanf                         ";
+  name[StatInt___isoc99_scanf]           = "  scanf                           ";
+  name[StatInt___isoc99_sscanf]          = "  sscanf                          ";
+  name[StatInt___isoc99_fscanf]          = "  fscanf                          ";
+  name[StatInt_on_exit]                  = "  on_exit                         ";
+  name[StatInt___cxa_atexit]             = "  __cxa_atexit                    ";
+  name[StatInt_localtime]                = "  localtime                       ";
+  name[StatInt_localtime_r]              = "  localtime_r                     ";
+  name[StatInt_gmtime]                   = "  gmtime                          ";
+  name[StatInt_gmtime_r]                 = "  gmtime_r                        ";
+  name[StatInt_ctime]                    = "  ctime                           ";
+  name[StatInt_ctime_r]                  = "  ctime_r                         ";
+  name[StatInt_asctime]                  = "  asctime                         ";
+  name[StatInt_asctime_r]                = "  asctime_r                       ";
+  name[StatInt_frexp]                    = "  frexp                           ";
+  name[StatInt_frexpf]                   = "  frexpf                          ";
+  name[StatInt_frexpl]                   = "  frexpl                          ";
+  name[StatInt_getpwnam]                 = "  getpwnam                        ";
+  name[StatInt_getpwuid]                 = "  getpwuid                        ";
+  name[StatInt_getgrnam]                 = "  getgrnam                        ";
+  name[StatInt_getgrgid]                 = "  getgrgid                        ";
+  name[StatInt_getpwnam_r]               = "  getpwnam_r                      ";
+  name[StatInt_getpwuid_r]               = "  getpwuid_r                      ";
+  name[StatInt_getgrnam_r]               = "  getgrnam_r                      ";
+  name[StatInt_getgrgid_r]               = "  getgrgid_r                      ";
+  name[StatInt_clock_getres]             = "  clock_getres                    ";
+  name[StatInt_clock_gettime]            = "  clock_gettime                   ";
+  name[StatInt_clock_settime]            = "  clock_settime                   ";
+  name[StatInt_getitimer]                = "  getitimer                       ";
+  name[StatInt_setitimer]                = "  setitimer                       ";
+  name[StatInt_time]                     = "  time                            ";
+  name[StatInt_glob]                     = "  glob                            ";
+  name[StatInt_glob64]                   = "  glob64                          ";
+  name[StatInt_wait]                     = "  wait                            ";
+  name[StatInt_waitid]                   = "  waitid                          ";
+  name[StatInt_waitpid]                  = "  waitpid                         ";
+  name[StatInt_wait3]                    = "  wait3                           ";
+  name[StatInt_wait4]                    = "  wait4                           ";
+  name[StatInt_inet_ntop]                = "  inet_ntop                       ";
+  name[StatInt_inet_pton]                = "  inet_pton                       ";
+  name[StatInt_getaddrinfo]              = "  getaddrinfo                     ";
+  name[StatInt_getsockname]              = "  getsockname                     ";
+  name[StatInt_gethostent]               = "  gethostent                      ";
+  name[StatInt_gethostbyname]            = "  gethostbyname                   ";
+  name[StatInt_gethostbyname2]           = "  gethostbyname2                  ";
+  name[StatInt_gethostbyaddr]            = "  gethostbyaddr                   ";
+  name[StatInt_gethostent_r]             = "  gethostent_r                    ";
+  name[StatInt_gethostbyname_r]          = "  gethostbyname_r                 ";
+  name[StatInt_gethostbyname2_r]         = "  gethostbyname2_r                ";
+  name[StatInt_gethostbyaddr_r]          = "  gethostbyaddr_r                 ";
+  name[StatInt_getsockopt]               = "  getsockopt                      ";
 
   name[StatAnnotation]                   = "Dynamic annotations               ";
   name[StatAnnotateHappensBefore]        = "  HappensBefore                   ";
@@ -251,6 +331,7 @@ void StatOutput(u64 *stat) {
   name[StatAnnotateMutexIsNotPHB]        = "  MutexIsNotPHB                   ";
   name[StatAnnotateCondVarWait]          = "  CondVarWait                     ";
   name[StatAnnotateRWLockCreate]         = "  RWLockCreate                    ";
+  name[StatAnnotateRWLockCreateStatic]   = "  StatAnnotateRWLockCreateStatic  ";
   name[StatAnnotateRWLockDestroy]        = "  RWLockDestroy                   ";
   name[StatAnnotateRWLockAcquired]       = "  RWLockAcquired                  ";
   name[StatAnnotateRWLockReleased]       = "  RWLockReleased                  ";
@@ -287,6 +368,7 @@ void StatOutput(u64 *stat) {
   name[StatMtxAnnotations]               = "  Annotations                     ";
   name[StatMtxMBlock]                    = "  MBlock                          ";
   name[StatMtxJavaMBlock]                = "  JavaMBlock                      ";
+  name[StatMtxFD]                        = "  FD                              ";
 
   Printf("Statistics:\n");
   for (int i = 0; i < StatCnt; i++)
diff --git a/lib/tsan/rtl/tsan_stat.h b/lib/tsan/rtl/tsan_stat.h
index 58c5f23af40b1..d5c8b43893942 100644
--- a/lib/tsan/rtl/tsan_stat.h
+++ b/lib/tsan/rtl/tsan_stat.h
@@ -27,6 +27,8 @@ enum StatType {
   StatMop8,
   StatMopSame,
   StatMopRange,
+  StatMopRodata,
+  StatMopRangeRodata,
   StatShadowProcessed,
   StatShadowZero,
   StatShadowNonZero,  // Derived.
@@ -102,6 +104,7 @@ enum StatType {
   StatInt_realloc,
   StatInt_free,
   StatInt_cfree,
+  StatInt_malloc_usable_size,
   StatInt_mmap,
   StatInt_mmap64,
   StatInt_munmap,
@@ -131,6 +134,8 @@ enum StatType {
   StatInt_strncmp,
   StatInt_strcpy,
   StatInt_strncpy,
+  StatInt_strcasecmp,
+  StatInt_strncasecmp,
   StatInt_strstr,
   StatInt_atexit,
   StatInt___cxa_guard_acquire,
@@ -169,6 +174,7 @@ enum StatType {
   StatInt_pthread_barrier_destroy,
   StatInt_pthread_barrier_wait,
   StatInt_pthread_once,
+  StatInt_pthread_getschedparam,
   StatInt_sem_init,
   StatInt_sem_destroy,
   StatInt_sem_wait,
@@ -176,6 +182,18 @@ enum StatType {
   StatInt_sem_timedwait,
   StatInt_sem_post,
   StatInt_sem_getvalue,
+  StatInt_stat,
+  StatInt___xstat,
+  StatInt_stat64,
+  StatInt___xstat64,
+  StatInt_lstat,
+  StatInt___lxstat,
+  StatInt_lstat64,
+  StatInt___lxstat64,
+  StatInt_fstat,
+  StatInt___fxstat,
+  StatInt_fstat64,
+  StatInt___fxstat64,
   StatInt_open,
   StatInt_open64,
   StatInt_creat,
@@ -190,12 +208,15 @@ enum StatType {
   StatInt_socket,
   StatInt_socketpair,
   StatInt_connect,
+  StatInt_bind,
+  StatInt_listen,
   StatInt_accept,
   StatInt_accept4,
   StatInt_epoll_create,
   StatInt_epoll_create1,
   StatInt_close,
   StatInt___close,
+  StatInt___res_iclose,
   StatInt_pipe,
   StatInt_pipe2,
   StatInt_read,
@@ -219,6 +240,8 @@ enum StatType {
   StatInt_fclose,
   StatInt_fread,
   StatInt_fwrite,
+  StatInt_fflush,
+  StatInt_abort,
   StatInt_puts,
   StatInt_rmdir,
   StatInt_opendir,
@@ -241,6 +264,59 @@ enum StatType {
   StatInt_scanf,
   StatInt_sscanf,
   StatInt_fscanf,
+  StatInt___isoc99_vscanf,
+  StatInt___isoc99_vsscanf,
+  StatInt___isoc99_vfscanf,
+  StatInt___isoc99_scanf,
+  StatInt___isoc99_sscanf,
+  StatInt___isoc99_fscanf,
+  StatInt_on_exit,
+  StatInt___cxa_atexit,
+  StatInt_localtime,
+  StatInt_localtime_r,
+  StatInt_gmtime,
+  StatInt_gmtime_r,
+  StatInt_ctime,
+  StatInt_ctime_r,
+  StatInt_asctime,
+  StatInt_asctime_r,
+  StatInt_frexp,
+  StatInt_frexpf,
+  StatInt_frexpl,
+  StatInt_getpwnam,
+  StatInt_getpwuid,
+  StatInt_getgrnam,
+  StatInt_getgrgid,
+  StatInt_getpwnam_r,
+  StatInt_getpwuid_r,
+  StatInt_getgrnam_r,
+  StatInt_getgrgid_r,
+  StatInt_clock_getres,
+  StatInt_clock_gettime,
+  StatInt_clock_settime,
+  StatInt_getitimer,
+  StatInt_setitimer,
+  StatInt_time,
+  StatInt_glob,
+  StatInt_glob64,
+  StatInt_wait,
+  StatInt_waitid,
+  StatInt_waitpid,
+  StatInt_wait3,
+  StatInt_wait4,
+  StatInt_inet_ntop,
+  StatInt_inet_pton,
+  StatInt_getaddrinfo,
+  StatInt_getsockname,
+  StatInt_gethostent,
+  StatInt_gethostbyname,
+  StatInt_gethostbyname2,
+  StatInt_gethostbyaddr,
+  StatInt_gethostent_r,
+  StatInt_gethostbyname_r,
+  StatInt_gethostbyname2_r,
+  StatInt_gethostbyaddr_r,
+  StatInt_getsockopt,
 
   // Dynamic annotations.
   StatAnnotation,
@@ -289,6 +365,7 @@ enum StatType {
   StatMtxAtExit,
   StatMtxMBlock,
   StatMtxJavaMBlock,
+  StatMtxFD,
 
   // This must be the last.
   StatCnt
diff --git a/lib/tsan/rtl/tsan_suppressions.cc b/lib/tsan/rtl/tsan_suppressions.cc
index 5316f6db6a0af..6c49355bed887 100644
--- a/lib/tsan/rtl/tsan_suppressions.cc
+++ b/lib/tsan/rtl/tsan_suppressions.cc
@@ -19,6 +19,13 @@
 #include "tsan_mman.h"
 #include "tsan_platform.h"
 
+// Can be overriden in frontend.
+#ifndef TSAN_GO
+extern "C" const char *WEAK __tsan_default_suppressions() {
+  return 0;
+}
+#endif
+
 namespace __tsan {
 
 static Suppression *g_suppressions;
@@ -31,12 +38,13 @@ static char *ReadFile(const char *filename) {
     internal_snprintf(tmp.data(), tmp.size(), "%s", filename);
   else
     internal_snprintf(tmp.data(), tmp.size(), "%s/%s", GetPwd(), filename);
-  fd_t fd = internal_open(tmp.data(), false);
-  if (fd == kInvalidFd) {
+  uptr openrv = OpenFile(tmp.data(), false);
+  if (internal_iserror(openrv)) {
     Printf("ThreadSanitizer: failed to open suppressions file '%s'\n",
                tmp.data());
     Die();
   }
+  fd_t fd = openrv;
   const uptr fsize = internal_filesize(fd);
   if (fsize == (uptr)-1) {
     Printf("ThreadSanitizer: failed to stat suppressions file '%s'\n",
@@ -80,8 +88,7 @@ bool SuppressionMatch(char *templ, const char *str) {
   return true;
 }
 
-Suppression *SuppressionParse(const char* supp) {
-  Suppression *head = 0;
+Suppression *SuppressionParse(Suppression *head, const char* supp) {
   const char *line = supp;
   while (line) {
     while (line[0] == ' ' || line[0] == '\t')
@@ -121,6 +128,7 @@ Suppression *SuppressionParse(const char* supp) {
       s->templ = (char*)internal_alloc(MBlockSuppression, end2 - line + 1);
       internal_memcpy(s->templ, line, end2 - line);
       s->templ[end2 - line] = 0;
+      s->hit_count = 0;
     }
     if (end[0] == 0)
       break;
@@ -130,11 +138,15 @@ Suppression *SuppressionParse(const char* supp) {
 }
 
 void InitializeSuppressions() {
-  char *supp = ReadFile(flags()->suppressions);
-  g_suppressions = SuppressionParse(supp);
+  const char *supp = ReadFile(flags()->suppressions);
+  g_suppressions = SuppressionParse(0, supp);
+#ifndef TSAN_GO
+  supp = __tsan_default_suppressions();
+  g_suppressions = SuppressionParse(g_suppressions, supp);
+#endif
 }
 
-uptr IsSuppressed(ReportType typ, const ReportStack *stack) {
+uptr IsSuppressed(ReportType typ, const ReportStack *stack, Suppression **sp) {
   if (g_suppressions == 0 || stack == 0)
     return 0;
   SuppressionType stype;
@@ -152,12 +164,41 @@ uptr IsSuppressed(ReportType typ, const ReportStack *stack) {
     for (Suppression *supp = g_suppressions; supp; supp = supp->next) {
       if (stype == supp->type &&
           (SuppressionMatch(supp->templ, frame->func) ||
-          SuppressionMatch(supp->templ, frame->file))) {
+           SuppressionMatch(supp->templ, frame->file) ||
+           SuppressionMatch(supp->templ, frame->module))) {
         DPrintf("ThreadSanitizer: matched suppression '%s'\n", supp->templ);
+        supp->hit_count++;
+        *sp = supp;
         return frame->pc;
       }
     }
   }
   return 0;
 }
+
+static const char *SuppTypeStr(SuppressionType t) {
+  switch (t) {
+  case SuppressionRace:   return "race";
+  case SuppressionMutex:  return "mutex";
+  case SuppressionThread: return "thread";
+  case SuppressionSignal: return "signal";
+  }
+  CHECK(0);
+  return "unknown";
+}
+
+void PrintMatchedSuppressions() {
+  int hit_count = 0;
+  for (Suppression *supp = g_suppressions; supp; supp = supp->next)
+    hit_count += supp->hit_count;
+  if (hit_count == 0)
+    return;
+  Printf("ThreadSanitizer: Matched %d suppressions (pid=%d):\n",
+      hit_count, (int)internal_getpid());
+  for (Suppression *supp = g_suppressions; supp; supp = supp->next) {
+    if (supp->hit_count == 0)
+      continue;
+    Printf("%d %s:%s\n", supp->hit_count, SuppTypeStr(supp->type), supp->templ);
+  }
+}
 }  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_suppressions.h b/lib/tsan/rtl/tsan_suppressions.h
index 61a4cca9d17ad..1c98363383dc1 100644
--- a/lib/tsan/rtl/tsan_suppressions.h
+++ b/lib/tsan/rtl/tsan_suppressions.h
@@ -17,10 +17,6 @@
 
 namespace __tsan {
 
-void InitializeSuppressions();
-void FinalizeSuppressions();
-uptr IsSuppressed(ReportType typ, const ReportStack *stack);
-
 // Exposed for testing.
 enum SuppressionType {
   SuppressionRace,
@@ -33,9 +29,14 @@ struct Suppression {
   Suppression *next;
   SuppressionType type;
   char *templ;
+  int hit_count;
 };
 
-Suppression *SuppressionParse(const char* supp);
+void InitializeSuppressions();
+void FinalizeSuppressions();
+void PrintMatchedSuppressions();
+uptr IsSuppressed(ReportType typ, const ReportStack *stack, Suppression **sp);
+Suppression *SuppressionParse(Suppression *head, const char* supp);
 bool SuppressionMatch(char *templ, const char *str);
 
 }  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_symbolize.cc b/lib/tsan/rtl/tsan_symbolize.cc
index 29dfe237ffd97..12226064f5a4b 100644
--- a/lib/tsan/rtl/tsan_symbolize.cc
+++ b/lib/tsan/rtl/tsan_symbolize.cc
@@ -18,9 +18,24 @@
 #include "sanitizer_common/sanitizer_symbolizer.h"
 #include "tsan_flags.h"
 #include "tsan_report.h"
+#include "tsan_rtl.h"
 
 namespace __tsan {
 
+struct ScopedInSymbolizer {
+  ScopedInSymbolizer() {
+    ThreadState *thr = cur_thread();
+    CHECK(!thr->in_symbolizer);
+    thr->in_symbolizer = true;
+  }
+
+  ~ScopedInSymbolizer() {
+    ThreadState *thr = cur_thread();
+    CHECK(thr->in_symbolizer);
+    thr->in_symbolizer = false;
+  }
+};
+
 ReportStack *NewReportStackEntry(uptr addr) {
   ReportStack *ent = (ReportStack*)internal_alloc(MBlockReportStack,
                                                   sizeof(ReportStack));
@@ -55,35 +70,36 @@ static ReportStack *NewReportStackEntry(const AddressInfo &info) {
 }
 
 ReportStack *SymbolizeCode(uptr addr) {
-  if (flags()->external_symbolizer_path[0]) {
-    static const uptr kMaxAddrFrames = 16;
-    InternalScopedBuffer<AddressInfo> addr_frames(kMaxAddrFrames);
-    for (uptr i = 0; i < kMaxAddrFrames; i++)
-      new(&addr_frames[i]) AddressInfo();
-    uptr addr_frames_num = __sanitizer::SymbolizeCode(addr, addr_frames.data(),
-                                                      kMaxAddrFrames);
-    if (addr_frames_num == 0)
-      return NewReportStackEntry(addr);
-    ReportStack *top = 0;
-    ReportStack *bottom = 0;
-    for (uptr i = 0; i < addr_frames_num; i++) {
-      ReportStack *cur_entry = NewReportStackEntry(addr_frames[i]);
-      CHECK(cur_entry);
-      addr_frames[i].Clear();
-      if (i == 0)
-        top = cur_entry;
-      else
-        bottom->next = cur_entry;
-      bottom = cur_entry;
-    }
-    return top;
+  if (!IsSymbolizerAvailable())
+    return SymbolizeCodeAddr2Line(addr);
+  ScopedInSymbolizer in_symbolizer;
+  static const uptr kMaxAddrFrames = 16;
+  InternalScopedBuffer<AddressInfo> addr_frames(kMaxAddrFrames);
+  for (uptr i = 0; i < kMaxAddrFrames; i++)
+    new(&addr_frames[i]) AddressInfo();
+  uptr addr_frames_num = __sanitizer::SymbolizeCode(addr, addr_frames.data(),
+                                                    kMaxAddrFrames);
+  if (addr_frames_num == 0)
+    return NewReportStackEntry(addr);
+  ReportStack *top = 0;
+  ReportStack *bottom = 0;
+  for (uptr i = 0; i < addr_frames_num; i++) {
+    ReportStack *cur_entry = NewReportStackEntry(addr_frames[i]);
+    CHECK(cur_entry);
+    addr_frames[i].Clear();
+    if (i == 0)
+      top = cur_entry;
+    else
+      bottom->next = cur_entry;
+    bottom = cur_entry;
   }
-  return SymbolizeCodeAddr2Line(addr);
+  return top;
 }
 
 ReportLocation *SymbolizeData(uptr addr) {
-  if (flags()->external_symbolizer_path[0] == 0)
+  if (!IsSymbolizerAvailable())
     return 0;
+  ScopedInSymbolizer in_symbolizer;
   DataInfo info;
   if (!__sanitizer::SymbolizeData(addr, &info))
     return 0;
@@ -100,4 +116,11 @@ ReportLocation *SymbolizeData(uptr addr) {
   return ent;
 }
 
+void SymbolizeFlush() {
+  if (!IsSymbolizerAvailable())
+    return;
+  ScopedInSymbolizer in_symbolizer;
+  __sanitizer::FlushSymbolizer();
+}
+
 }  // namespace __tsan
diff --git a/lib/tsan/rtl/tsan_symbolize.h b/lib/tsan/rtl/tsan_symbolize.h
index 29193043cd709..7bc6123df57de 100644
--- a/lib/tsan/rtl/tsan_symbolize.h
+++ b/lib/tsan/rtl/tsan_symbolize.h
@@ -20,6 +20,7 @@ namespace __tsan {
 
 ReportStack *SymbolizeCode(uptr addr);
 ReportLocation *SymbolizeData(uptr addr);
+void SymbolizeFlush();
 
 ReportStack *SymbolizeCodeAddr2Line(uptr addr);
 
diff --git a/lib/tsan/rtl/tsan_symbolize_addr2line_linux.cc b/lib/tsan/rtl/tsan_symbolize_addr2line_linux.cc
index 76926e2b5aaf1..47f9e1fbf4189 100644
--- a/lib/tsan/rtl/tsan_symbolize_addr2line_linux.cc
+++ b/lib/tsan/rtl/tsan_symbolize_addr2line_linux.cc
@@ -87,7 +87,8 @@ static int dl_iterate_phdr_cb(dl_phdr_info *info, size_t size, void *arg) {
   DlIteratePhdrCtx *ctx = (DlIteratePhdrCtx*)arg;
   InternalScopedBuffer<char> tmp(128);
   if (ctx->is_first) {
-    internal_snprintf(tmp.data(), tmp.size(), "/proc/%d/exe", GetPid());
+    internal_snprintf(tmp.data(), tmp.size(), "/proc/%d/exe",
+                      (int)internal_getpid());
     info->dlpi_name = tmp.data();
   }
   ctx->is_first = false;
diff --git a/lib/tsan/rtl/tsan_sync.cc b/lib/tsan/rtl/tsan_sync.cc
index b25346ef344f9..c6ddcdb374260 100644
--- a/lib/tsan/rtl/tsan_sync.cc
+++ b/lib/tsan/rtl/tsan_sync.cc
@@ -63,7 +63,7 @@ SyncVar* SyncTab::Create(ThreadState *thr, uptr pc, uptr addr) {
   const u64 uid = atomic_fetch_add(&uid_gen_, 1, memory_order_relaxed);
   SyncVar *res = new(mem) SyncVar(addr, uid);
 #ifndef TSAN_GO
-  res->creation_stack.ObtainCurrent(thr, pc);
+  res->creation_stack_id = CurrentStackId(thr, pc);
 #endif
   return res;
 }
@@ -82,9 +82,10 @@ SyncVar* SyncTab::GetAndLock(ThreadState *thr, uptr pc,
   // the hashmap anyway.
   if (PrimaryAllocator::PointerIsMine((void*)addr)) {
     MBlock *b = user_mblock(thr, (void*)addr);
-    Lock l(&b->mtx);
+    CHECK_NE(b, 0);
+    MBlock::ScopedLock l(b);
     SyncVar *res = 0;
-    for (res = b->head; res; res = res->next) {
+    for (res = b->ListHead(); res; res = res->next) {
       if (res->addr == addr)
         break;
     }
@@ -92,8 +93,7 @@ SyncVar* SyncTab::GetAndLock(ThreadState *thr, uptr pc,
       if (!create)
         return 0;
       res = Create(thr, pc, addr);
-      res->next = b->head;
-      b->head = res;
+      b->ListPush(res);
     }
     if (write_lock)
       res->mtx.Lock();
@@ -147,27 +147,37 @@ SyncVar* SyncTab::GetAndRemove(ThreadState *thr, uptr pc, uptr addr) {
   }
   if (PrimaryAllocator::PointerIsMine((void*)addr)) {
     MBlock *b = user_mblock(thr, (void*)addr);
+    CHECK_NE(b, 0);
     SyncVar *res = 0;
     {
-      Lock l(&b->mtx);
-      SyncVar **prev = &b->head;
-      res = *prev;
-      while (res) {
+      MBlock::ScopedLock l(b);
+      res = b->ListHead();
+      if (res) {
         if (res->addr == addr) {
           if (res->is_linker_init)
             return 0;
-          *prev = res->next;
-          break;
+          b->ListPop();
+        } else {
+          SyncVar **prev = &res->next;
+          res = *prev;
+          while (res) {
+            if (res->addr == addr) {
+              if (res->is_linker_init)
+                return 0;
+              *prev = res->next;
+              break;
+            }
+            prev = &res->next;
+            res = *prev;
+          }
+        }
+        if (res) {
+          StatInc(thr, StatSyncDestroyed);
+          res->mtx.Lock();
+          res->mtx.Unlock();
         }
-        prev = &res->next;
-        res = *prev;
       }
     }
-    if (res) {
-      StatInc(thr, StatSyncDestroyed);
-      res->mtx.Lock();
-      res->mtx.Unlock();
-    }
     return res;
   }
 #endif
@@ -197,26 +207,6 @@ SyncVar* SyncTab::GetAndRemove(ThreadState *thr, uptr pc, uptr addr) {
   return res;
 }
 
-uptr SyncVar::GetMemoryConsumption() {
-  return sizeof(*this)
-      + clock.size() * sizeof(u64)
-      + read_clock.size() * sizeof(u64)
-      + creation_stack.Size() * sizeof(uptr);
-}
-
-uptr SyncTab::GetMemoryConsumption(uptr *nsync) {
-  uptr mem = 0;
-  for (int i = 0; i < kPartCount; i++) {
-    Part *p = &tab_[i];
-    Lock l(&p->mtx);
-    for (SyncVar *s = p->val; s; s = s->next) {
-      *nsync += 1;
-      mem += s->GetMemoryConsumption();
-    }
-  }
-  return mem;
-}
-
 int SyncTab::PartIdx(uptr addr) {
   return (addr >> 3) % kPartCount;
 }
diff --git a/lib/tsan/rtl/tsan_sync.h b/lib/tsan/rtl/tsan_sync.h
index 77749e22ffc2a..823af543f590b 100644
--- a/lib/tsan/rtl/tsan_sync.h
+++ b/lib/tsan/rtl/tsan_sync.h
@@ -59,7 +59,7 @@ struct SyncVar {
   const u64 uid;  // Globally unique id.
   SyncClock clock;
   SyncClock read_clock;  // Used for rw mutexes only.
-  StackTrace creation_stack;
+  u32 creation_stack_id;
   int owner_tid;  // Set only by exclusive owners.
   u64 last_lock;
   int recursion;
diff --git a/lib/tsan/rtl/tsan_update_shadow_word_inl.h b/lib/tsan/rtl/tsan_update_shadow_word_inl.h
index 2c435556abb20..e7c036c5dea87 100644
--- a/lib/tsan/rtl/tsan_update_shadow_word_inl.h
+++ b/lib/tsan/rtl/tsan_update_shadow_word_inl.h
@@ -34,7 +34,7 @@ do {
     if (Shadow::TidsAreEqual(old, cur)) {
       StatInc(thr, StatShadowSameThread);
       if (OldIsInSameSynchEpoch(old, thr)) {
-        if (OldIsRWNotWeaker(old, kAccessIsWrite)) {
+        if (old.IsRWNotWeaker(kAccessIsWrite, kIsAtomic)) {
           // found a slot that holds effectively the same info
           // (that is, same tid, same sync epoch and same size)
           StatInc(thr, StatMopSame);
@@ -43,7 +43,7 @@ do {
         StoreIfNotYetStored(sp, &store_word);
         break;
       }
-      if (OldIsRWWeakerOrEqual(old, kAccessIsWrite))
+      if (old.IsRWWeakerOrEqual(kAccessIsWrite, kIsAtomic))
         StoreIfNotYetStored(sp, &store_word);
       break;
     }
@@ -52,25 +52,23 @@ do {
       StoreIfNotYetStored(sp, &store_word);
       break;
     }
-    if (BothReads(old, kAccessIsWrite))
+    if (old.IsBothReadsOrAtomic(kAccessIsWrite, kIsAtomic))
       break;
     goto RACE;
   }
-
   // Do the memory access intersect?
-  if (Shadow::TwoRangesIntersect(old, cur, kAccessSize)) {
+  // In Go all memory accesses are 1 byte, so there can be no intersections.
+  if (kCppMode && Shadow::TwoRangesIntersect(old, cur, kAccessSize)) {
     StatInc(thr, StatShadowIntersect);
     if (Shadow::TidsAreEqual(old, cur)) {
       StatInc(thr, StatShadowSameThread);
       break;
     }
     StatInc(thr, StatShadowAnotherThread);
-    if (HappensBefore(old, thr))
+    if (old.IsBothReadsOrAtomic(kAccessIsWrite, kIsAtomic))
       break;
-
-    if (BothReads(old, kAccessIsWrite))
+    if (HappensBefore(old, thr))
       break;
-
     goto RACE;
   }
   // The accesses do not intersect.
diff --git a/lib/tsan/rtl/tsan_vector.h b/lib/tsan/rtl/tsan_vector.h
index d41063df3de5b..fa236b1f1e44d 100644
--- a/lib/tsan/rtl/tsan_vector.h
+++ b/lib/tsan/rtl/tsan_vector.h
@@ -64,6 +64,11 @@ class Vector {
     return &end_[-1];
   }
 
+  void PopBack() {
+    DCHECK_GT(end_, begin_);
+    end_--;
+  }
+
   void Resize(uptr size) {
     uptr old_size = Size();
     EnsureSize(size);
@@ -105,6 +110,6 @@ class Vector {
   Vector(const Vector&);
   void operator=(const Vector&);
 };
-}
+}  // namespace __tsan
 
 #endif  // #ifndef TSAN_VECTOR_H
diff --git a/lib/tsan/tests/CMakeLists.txt b/lib/tsan/tests/CMakeLists.txt
index 0fcc6b2b1c8f3..7cc079f3d27af 100644
--- a/lib/tsan/tests/CMakeLists.txt
+++ b/lib/tsan/tests/CMakeLists.txt
@@ -12,9 +12,9 @@ function(add_tsan_unittest testname)
     add_unittest(TsanUnitTests ${testname} ${ARGN})
     # Link with TSan runtime.
     target_link_libraries(${testname} clang_rt.tsan-x86_64)
-    # Build tests with PIE and debug info.
-    set_property(TARGET ${testname} APPEND_STRING
-      PROPERTY COMPILE_FLAGS " -fPIE -g")
+    # Compile tests with the same flags as TSan runtime.
+    set_target_compile_flags(${testname} ${TSAN_CFLAGS})
+    # Link tests with -pie.
     set_property(TARGET ${testname} APPEND_STRING
       PROPERTY LINK_FLAGS " -pie")
   endif()
diff --git a/lib/tsan/tests/rtl/tsan_test_util_linux.cc b/lib/tsan/tests/rtl/tsan_test_util_linux.cc
index dce8db90de70b..a2601486a2e1f 100644
--- a/lib/tsan/tests/rtl/tsan_test_util_linux.cc
+++ b/lib/tsan/tests/rtl/tsan_test_util_linux.cc
@@ -73,7 +73,7 @@ bool OnReport(const ReportDesc *rep, bool suppressed) {
   expect_report_reported = true;
   return true;
 }
-}
+}  // namespace __tsan
 
 static void* allocate_addr(int size, int offset_from_aligned = 0) {
   static uintptr_t foo;
diff --git a/lib/tsan/tests/unit/CMakeLists.txt b/lib/tsan/tests/unit/CMakeLists.txt
index 52ebdb8269393..b25a56d8d55c8 100644
--- a/lib/tsan/tests/unit/CMakeLists.txt
+++ b/lib/tsan/tests/unit/CMakeLists.txt
@@ -3,7 +3,6 @@ set(TSAN_UNIT_TESTS
   tsan_flags_test.cc
   tsan_mman_test.cc
   tsan_mutex_test.cc
-  tsan_platform_test.cc
   tsan_shadow_test.cc
   tsan_stack_test.cc
   tsan_suppressions_test.cc
diff --git a/lib/tsan/tests/unit/tsan_mman_test.cc b/lib/tsan/tests/unit/tsan_mman_test.cc
index 1a9a88f606fcc..0961d2b75d11c 100644
--- a/lib/tsan/tests/unit/tsan_mman_test.cc
+++ b/lib/tsan/tests/unit/tsan_mman_test.cc
@@ -10,10 +10,21 @@
 // This file is a part of ThreadSanitizer (TSan), a race detector.
 //
 //===----------------------------------------------------------------------===//
+#include <limits>
 #include "tsan_mman.h"
 #include "tsan_rtl.h"
 #include "gtest/gtest.h"
 
+extern "C" {
+uptr __tsan_get_current_allocated_bytes();
+uptr __tsan_get_heap_size();
+uptr __tsan_get_free_bytes();
+uptr __tsan_get_unmapped_bytes();
+uptr __tsan_get_estimated_allocated_size(uptr size);
+bool __tsan_get_ownership(void *p);
+uptr __tsan_get_allocated_size(void *p);
+}
+
 namespace __tsan {
 
 TEST(Mman, Internal) {
@@ -44,10 +55,10 @@ TEST(Mman, User) {
   EXPECT_NE(p2, p);
   MBlock *b = user_mblock(thr, p);
   EXPECT_NE(b, (MBlock*)0);
-  EXPECT_EQ(b->size, (uptr)10);
+  EXPECT_EQ(b->Size(), (uptr)10);
   MBlock *b2 = user_mblock(thr, p2);
   EXPECT_NE(b2, (MBlock*)0);
-  EXPECT_EQ(b2->size, (uptr)20);
+  EXPECT_EQ(b2->Size(), (uptr)20);
   for (int i = 0; i < 10; i++) {
     p[i] = 42;
     EXPECT_EQ(b, user_mblock(thr, p + i));
@@ -106,4 +117,55 @@ TEST(Mman, UserRealloc) {
   }
 }
 
+TEST(Mman, UsableSize) {
+  ScopedInRtl in_rtl;
+  ThreadState *thr = cur_thread();
+  uptr pc = 0;
+  char *p = (char*)user_alloc(thr, pc, 10);
+  char *p2 = (char*)user_alloc(thr, pc, 20);
+  EXPECT_EQ(0U, user_alloc_usable_size(thr, pc, NULL));
+  EXPECT_EQ(10U, user_alloc_usable_size(thr, pc, p));
+  EXPECT_EQ(20U, user_alloc_usable_size(thr, pc, p2));
+  user_free(thr, pc, p);
+  user_free(thr, pc, p2);
+}
+
+TEST(Mman, Stats) {
+  ScopedInRtl in_rtl;
+  ThreadState *thr = cur_thread();
+
+  uptr alloc0 = __tsan_get_current_allocated_bytes();
+  uptr heap0 = __tsan_get_heap_size();
+  uptr free0 = __tsan_get_free_bytes();
+  uptr unmapped0 = __tsan_get_unmapped_bytes();
+
+  EXPECT_EQ(__tsan_get_estimated_allocated_size(10), (uptr)10);
+  EXPECT_EQ(__tsan_get_estimated_allocated_size(20), (uptr)20);
+  EXPECT_EQ(__tsan_get_estimated_allocated_size(100), (uptr)100);
+
+  char *p = (char*)user_alloc(thr, 0, 10);
+  EXPECT_EQ(__tsan_get_ownership(p), true);
+  EXPECT_EQ(__tsan_get_allocated_size(p), (uptr)10);
+
+  EXPECT_EQ(__tsan_get_current_allocated_bytes(), alloc0 + 16);
+  EXPECT_GE(__tsan_get_heap_size(), heap0);
+  EXPECT_EQ(__tsan_get_free_bytes(), free0);
+  EXPECT_EQ(__tsan_get_unmapped_bytes(), unmapped0);
+
+  user_free(thr, 0, p);
+
+  EXPECT_EQ(__tsan_get_current_allocated_bytes(), alloc0);
+  EXPECT_GE(__tsan_get_heap_size(), heap0);
+  EXPECT_EQ(__tsan_get_free_bytes(), free0);
+  EXPECT_EQ(__tsan_get_unmapped_bytes(), unmapped0);
+}
+
+TEST(Mman, CallocOverflow) {
+  size_t kArraySize = 4096;
+  volatile size_t kMaxSizeT = std::numeric_limits<size_t>::max();
+  volatile size_t kArraySize2 = kMaxSizeT / kArraySize + 10;
+  volatile void *p = calloc(kArraySize, kArraySize2);  // Should return 0.
+  EXPECT_EQ(0L, p);
+}
+
 }  // namespace __tsan
diff --git a/lib/tsan/tests/unit/tsan_mutexset_test.cc b/lib/tsan/tests/unit/tsan_mutexset_test.cc
index da1ae2e49e0c9..335a7748cc1a9 100644
--- a/lib/tsan/tests/unit/tsan_mutexset_test.cc
+++ b/lib/tsan/tests/unit/tsan_mutexset_test.cc
@@ -115,7 +115,8 @@ TEST(MutexSet, Overflow) {
   EXPECT_EQ(mset.Size(), MutexSet::kMaxSize);
   for (uptr i = 0; i < MutexSet::kMaxSize; i++) {
     if (i == 0)
-      Expect(mset, i, 63, true, 64, 2);
+      Expect(mset, i, MutexSet::kMaxSize - 1,
+             true, MutexSet::kMaxSize, 2);
     else if (i == MutexSet::kMaxSize - 1)
       Expect(mset, i, 100, true, 200, 1);
     else
diff --git a/lib/tsan/tests/unit/tsan_platform_test.cc b/lib/tsan/tests/unit/tsan_platform_test.cc
deleted file mode 100644
index b43dbb4e4ff35..0000000000000
--- a/lib/tsan/tests/unit/tsan_platform_test.cc
+++ /dev/null
@@ -1,89 +0,0 @@
-//===-- tsan_platform_test.cc ---------------------------------------------===//
-//
-//                     The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-//
-// This file is a part of ThreadSanitizer (TSan), a race detector.
-//
-//===----------------------------------------------------------------------===//
-#include "sanitizer_common/sanitizer_libc.h"
-#include "tsan_platform.h"
-#include "tsan_rtl.h"
-#include "gtest/gtest.h"
-
-namespace __tsan {
-
-static void TestThreadInfo(bool main) {
-  ScopedInRtl in_rtl;
-  uptr stk_addr = 0;
-  uptr stk_size = 0;
-  uptr tls_addr = 0;
-  uptr tls_size = 0;
-  GetThreadStackAndTls(main, &stk_addr, &stk_size, &tls_addr, &tls_size);
-  // Printf("stk=%zx-%zx(%zu)\n", stk_addr, stk_addr + stk_size, stk_size);
-  // Printf("tls=%zx-%zx(%zu)\n", tls_addr, tls_addr + tls_size, tls_size);
-
-  int stack_var;
-  EXPECT_NE(stk_addr, (uptr)0);
-  EXPECT_NE(stk_size, (uptr)0);
-  EXPECT_GT((uptr)&stack_var, stk_addr);
-  EXPECT_LT((uptr)&stack_var, stk_addr + stk_size);
-
-  static __thread int thread_var;
-  EXPECT_NE(tls_addr, (uptr)0);
-  EXPECT_NE(tls_size, (uptr)0);
-  EXPECT_GT((uptr)&thread_var, tls_addr);
-  EXPECT_LT((uptr)&thread_var, tls_addr + tls_size);
-
-  // Ensure that tls and stack do not intersect.
-  uptr tls_end = tls_addr + tls_size;
-  EXPECT_TRUE(tls_addr < stk_addr || tls_addr >= stk_addr + stk_size);
-  EXPECT_TRUE(tls_end  < stk_addr || tls_end  >=  stk_addr + stk_size);
-  EXPECT_TRUE((tls_addr < stk_addr) == (tls_end  < stk_addr));
-}
-
-static void *WorkerThread(void *arg) {
-  TestThreadInfo(false);
-  return 0;
-}
-
-TEST(Platform, ThreadInfoMain) {
-  TestThreadInfo(true);
-}
-
-TEST(Platform, ThreadInfoWorker) {
-  pthread_t t;
-  pthread_create(&t, 0, WorkerThread, 0);
-  pthread_join(t, 0);
-}
-
-TEST(Platform, FileOps) {
-  const char *str1 = "qwerty";
-  uptr len1 = internal_strlen(str1);
-  const char *str2 = "zxcv";
-  uptr len2 = internal_strlen(str2);
-
-  fd_t fd = internal_open("./tsan_test.tmp", true);
-  EXPECT_NE(fd, kInvalidFd);
-  EXPECT_EQ(len1, internal_write(fd, str1, len1));
-  EXPECT_EQ(len2, internal_write(fd, str2, len2));
-  internal_close(fd);
-
-  fd = internal_open("./tsan_test.tmp", false);
-  EXPECT_NE(fd, kInvalidFd);
-  EXPECT_EQ(len1 + len2, internal_filesize(fd));
-  char buf[64] = {};
-  EXPECT_EQ(len1, internal_read(fd, buf, len1));
-  EXPECT_EQ(0, internal_memcmp(buf, str1, len1));
-  EXPECT_EQ((char)0, buf[len1 + 1]);
-  internal_memset(buf, 0, len1);
-  EXPECT_EQ(len2, internal_read(fd, buf, len2));
-  EXPECT_EQ(0, internal_memcmp(buf, str2, len2));
-  internal_close(fd);
-}
-
-}  // namespace __tsan
diff --git a/lib/tsan/tests/unit/tsan_shadow_test.cc b/lib/tsan/tests/unit/tsan_shadow_test.cc
index fa9c982c0f6da..17b17977bf86b 100644
--- a/lib/tsan/tests/unit/tsan_shadow_test.cc
+++ b/lib/tsan/tests/unit/tsan_shadow_test.cc
@@ -25,7 +25,7 @@ TEST(Shadow, FastState) {
   EXPECT_EQ(s.GetHistorySize(), 0);
   EXPECT_EQ(s.addr0(), (u64)0);
   EXPECT_EQ(s.size(), (u64)1);
-  EXPECT_EQ(s.is_write(), false);
+  EXPECT_EQ(s.IsWrite(), true);
 
   s.IncrementEpoch();
   EXPECT_EQ(s.epoch(), (u64)23);
diff --git a/lib/tsan/tests/unit/tsan_suppressions_test.cc b/lib/tsan/tests/unit/tsan_suppressions_test.cc
index e1e0c12c004c1..decfa3214d23b 100644
--- a/lib/tsan/tests/unit/tsan_suppressions_test.cc
+++ b/lib/tsan/tests/unit/tsan_suppressions_test.cc
@@ -20,7 +20,7 @@ namespace __tsan {
 
 TEST(Suppressions, Parse) {
   ScopedInRtl in_rtl;
-  Suppression *supp0 = SuppressionParse(
+  Suppression *supp0 = SuppressionParse(0,
     "race:foo\n"
     " 	race:bar\n"  // NOLINT
     "race:baz	 \n"  // NOLINT
@@ -45,7 +45,7 @@ TEST(Suppressions, Parse) {
 
 TEST(Suppressions, Parse2) {
   ScopedInRtl in_rtl;
-  Suppression *supp0 = SuppressionParse(
+  Suppression *supp0 = SuppressionParse(0,
     "  	# first line comment\n"  // NOLINT
     " 	race:bar 	\n"  // NOLINT
     "race:baz* *baz\n"
@@ -64,7 +64,7 @@ TEST(Suppressions, Parse2) {
 
 TEST(Suppressions, Parse3) {
   ScopedInRtl in_rtl;
-  Suppression *supp0 = SuppressionParse(
+  Suppression *supp0 = SuppressionParse(0,
     "# last suppression w/o line-feed\n"
     "race:foo\n"
     "race:bar"
@@ -81,7 +81,7 @@ TEST(Suppressions, Parse3) {
 
 TEST(Suppressions, ParseType) {
   ScopedInRtl in_rtl;
-  Suppression *supp0 = SuppressionParse(
+  Suppression *supp0 = SuppressionParse(0,
     "race:foo\n"
     "thread:bar\n"
     "mutex:baz\n"