Index: compiler-rt/trunk/lib/hwasan/CMakeLists.txt
===================================================================
--- compiler-rt/trunk/lib/hwasan/CMakeLists.txt
+++ compiler-rt/trunk/lib/hwasan/CMakeLists.txt
@@ -10,6 +10,7 @@
   hwasan_poisoning.cc
   hwasan_report.cc
   hwasan_thread.cc
+  hwasan_thread_list.cc
   )
 
 set(HWASAN_RTL_CXX_SOURCES
@@ -25,8 +26,9 @@
   hwasan_mapping.h
   hwasan_poisoning.h
   hwasan_report.h
-  hwasan_thread.h)
-
+  hwasan_thread.h
+  hwasan_thread_list.h
+  )
 
 set(HWASAN_DEFINITIONS)
 append_list_if(COMPILER_RT_HWASAN_WITH_INTERCEPTORS HWASAN_WITH_INTERCEPTORS=1 HWASAN_DEFINITIONS)
Index: compiler-rt/trunk/lib/hwasan/hwasan.h
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan.h
+++ compiler-rt/trunk/lib/hwasan/hwasan.h
@@ -41,6 +41,10 @@
 const unsigned kAddressTagShift = 56;
 const uptr kAddressTagMask = 0xFFUL << kAddressTagShift;
 
+// Minimal alignment of the shadow base address. Determines the space available
+// for threads and stack histories. This is an ABI constant.
+const unsigned kShadowBaseAlignment = 32;
+
 static inline tag_t GetTagFromPointer(uptr p) {
   return p >> kAddressTagShift;
 }
@@ -66,6 +70,7 @@
 
 bool ProtectRange(uptr beg, uptr end);
 bool InitShadow();
+void InitThreads();
 void MadviseShadow();
 char *GetProcSelfMaps();
 void InitializeInterceptors();
@@ -142,6 +147,7 @@
 };
 
 void HwasanTSDInit();
+void HwasanTSDThreadInit();
 
 void HwasanOnDeadlySignal(int signo, void *info, void *context);
 
Index: compiler-rt/trunk/lib/hwasan/hwasan.cc
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan.cc
+++ compiler-rt/trunk/lib/hwasan/hwasan.cc
@@ -17,6 +17,7 @@
 #include "hwasan_poisoning.h"
 #include "hwasan_report.h"
 #include "hwasan_thread.h"
+#include "hwasan_thread_list.h"
 #include "sanitizer_common/sanitizer_atomic.h"
 #include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_flags.h"
@@ -174,7 +175,8 @@
 static constexpr uptr kMemoryUsageBufferSize = 4096;
 
 static void HwasanFormatMemoryUsage(InternalScopedString &s) {
-  auto thread_stats = Thread::GetThreadStats();
+  HwasanThreadList &thread_list = hwasanThreadList();
+  auto thread_stats = thread_list.GetThreadStats();
   auto *sds = StackDepotGetStats();
   AllocatorStatCounters asc;
   GetAllocatorStats(asc);
@@ -184,7 +186,7 @@
       " heap: %zd",
       internal_getpid(), GetRSS(), thread_stats.n_live_threads,
       thread_stats.total_stack_size,
-      thread_stats.n_live_threads * Thread::MemoryUsedPerThread(),
+      thread_stats.n_live_threads * thread_list.MemoryUsedPerThread(),
       sds->allocated, sds->n_uniq_ids, asc[AllocatorStatMapped]);
 }
 
@@ -253,7 +255,12 @@
   __sanitizer_set_report_path(common_flags()->log_path);
 
   DisableCoreDumperIfNecessary();
+
   __hwasan_shadow_init();
+
+  InitThreads();
+  hwasanThreadList().CreateCurrentThread();
+
   MadviseShadow();
 
   // This may call libc -> needs initialized shadow.
@@ -268,11 +275,10 @@
   InitializeCoverage(common_flags()->coverage, common_flags()->coverage_dir);
 
   HwasanTSDInit();
+  HwasanTSDThreadInit();
 
   HwasanAllocatorInit();
 
-  Thread::Create();
-
 #if HWASAN_CONTAINS_UBSAN
   __ubsan::InitAsPlugin();
 #endif
Index: compiler-rt/trunk/lib/hwasan/hwasan_dynamic_shadow.cc
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_dynamic_shadow.cc
+++ compiler-rt/trunk/lib/hwasan/hwasan_dynamic_shadow.cc
@@ -13,6 +13,7 @@
 ///
 //===----------------------------------------------------------------------===//
 
+#include "hwasan.h"
 #include "hwasan_dynamic_shadow.h"
 #include "hwasan_mapping.h"
 #include "sanitizer_common/sanitizer_common.h"
@@ -35,12 +36,16 @@
   }
 }
 
-// Returns an address aligned to 8 pages, such that one page on the left and
-// shadow_size_bytes bytes on the right of it are mapped r/o.
+// Returns an address aligned to kShadowBaseAlignment, such that
+// 2**kShadowBaseAlingment on the left and shadow_size_bytes bytes on the right
+// of it are mapped no access.
 static uptr MapDynamicShadow(uptr shadow_size_bytes) {
   const uptr granularity = GetMmapGranularity();
-  const uptr alignment = granularity << kShadowScale;
-  const uptr left_padding = granularity;
+  const uptr min_alignment = granularity << kShadowScale;
+  const uptr alignment = 1ULL << kShadowBaseAlignment;
+  CHECK_GE(alignment, min_alignment);
+
+  const uptr left_padding = 1ULL << kShadowBaseAlignment;
   const uptr shadow_size =
       RoundUpTo(shadow_size_bytes, granularity);
   const uptr map_size = shadow_size + left_padding + alignment;
Index: compiler-rt/trunk/lib/hwasan/hwasan_flags.inc
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_flags.inc
+++ compiler-rt/trunk/lib/hwasan/hwasan_flags.inc
@@ -51,3 +51,7 @@
           "to find bugs.")
 HWASAN_FLAG(bool, export_memory_stats, true,
             "Export up-to-date memory stats through /proc")
+HWASAN_FLAG(int, stack_history_size, 1024,
+            "The number of stack frames remembered per thread. "
+            "Affects the quality of stack-related reports, but not the ability "
+            "to find bugs.")
Index: compiler-rt/trunk/lib/hwasan/hwasan_linux.cc
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_linux.cc
+++ compiler-rt/trunk/lib/hwasan/hwasan_linux.cc
@@ -22,6 +22,7 @@
 #include "hwasan_mapping.h"
 #include "hwasan_report.h"
 #include "hwasan_thread.h"
+#include "hwasan_thread_list.h"
 
 #include <elf.h>
 #include <link.h>
@@ -37,6 +38,10 @@
 #include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_procmaps.h"
 
+#if HWASAN_WITH_INTERCEPTORS && !SANITIZER_ANDROID
+THREADLOCAL uptr __hwasan_tls;
+#endif
+
 namespace __hwasan {
 
 static void ReserveShadowMemoryRange(uptr beg, uptr end, const char *name) {
@@ -179,6 +184,20 @@
   return true;
 }
 
+void InitThreads() {
+  CHECK(__hwasan_shadow_memory_dynamic_address);
+  uptr guard_page_size = GetMmapGranularity();
+  uptr thread_space_start =
+      __hwasan_shadow_memory_dynamic_address - (1ULL << kShadowBaseAlignment);
+  uptr thread_space_end =
+      __hwasan_shadow_memory_dynamic_address - guard_page_size;
+  ReserveShadowMemoryRange(thread_space_start, thread_space_end - 1,
+                           "hwasan threads");
+  ProtectGap(thread_space_end,
+             __hwasan_shadow_memory_dynamic_address - thread_space_end);
+  InitThreadList(thread_space_start, thread_space_end - thread_space_start);
+}
+
 static void MadviseShadowRegion(uptr beg, uptr end) {
   uptr size = end - beg + 1;
   if (common_flags()->no_huge_pages_for_shadow)
@@ -214,7 +233,7 @@
 // ---------------------- TSD ---------------- {{{1
 
 extern "C" void __hwasan_thread_enter() {
-  Thread::Create();
+  hwasanThreadList().CreateCurrentThread();
 }
 
 extern "C" void __hwasan_thread_exit() {
@@ -222,21 +241,25 @@
   // Make sure that signal handler can not see a stale current thread pointer.
   atomic_signal_fence(memory_order_seq_cst);
   if (t)
-    t->Destroy();
+    hwasanThreadList().ReleaseThread(t);
 }
 
 #if HWASAN_WITH_INTERCEPTORS
 static pthread_key_t tsd_key;
 static bool tsd_key_inited = false;
 
+void HwasanTSDThreadInit() {
+  if (tsd_key_inited)
+    CHECK_EQ(0, pthread_setspecific(tsd_key,
+                                    (void *)GetPthreadDestructorIterations()));
+}
+
 void HwasanTSDDtor(void *tsd) {
-  Thread *t = (Thread*)tsd;
-  if (t->destructor_iterations_ > 1) {
-    t->destructor_iterations_--;
-    CHECK_EQ(0, pthread_setspecific(tsd_key, tsd));
+  uptr iterations = (uptr)tsd;
+  if (iterations > 1) {
+    CHECK_EQ(0, pthread_setspecific(tsd_key, (void *)(iterations - 1)));
     return;
   }
-  t->Destroy();
   __hwasan_thread_exit();
 }
 
@@ -245,31 +268,26 @@
   tsd_key_inited = true;
   CHECK_EQ(0, pthread_key_create(&tsd_key, HwasanTSDDtor));
 }
-
-Thread *GetCurrentThread() {
-  return (Thread *)pthread_getspecific(tsd_key);
-}
-
-void SetCurrentThread(Thread *t) {
-  // Make sure that HwasanTSDDtor gets called at the end.
-  CHECK(tsd_key_inited);
-  // Make sure we do not reset the current Thread.
-  CHECK_EQ(0, pthread_getspecific(tsd_key));
-  pthread_setspecific(tsd_key, (void *)t);
-}
-#elif SANITIZER_ANDROID
+#else
 void HwasanTSDInit() {}
-Thread *GetCurrentThread() {
-  return (Thread*)*get_android_tls_ptr();
-}
+void HwasanTSDThreadInit() {}
+#endif
 
-void SetCurrentThread(Thread *t) {
-  *get_android_tls_ptr() = (uptr)t;
+#if SANITIZER_ANDROID
+uptr *GetCurrentThreadLongPtr() {
+  return (uptr *)get_android_tls_ptr();
 }
 #else
-#error unsupported configuration !HWASAN_WITH_INTERCEPTORS && !SANITIZER_ANDROID
+uptr *GetCurrentThreadLongPtr() {
+  return &__hwasan_tls;
+}
 #endif
 
+Thread *GetCurrentThread() {
+  auto *R = (StackAllocationsRingBuffer*)GetCurrentThreadLongPtr();
+  return hwasanThreadList().GetThreadByBufferAddress((uptr)(R->Next()));
+}
+
 struct AccessInfo {
   uptr addr;
   uptr size;
Index: compiler-rt/trunk/lib/hwasan/hwasan_report.cc
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_report.cc
+++ compiler-rt/trunk/lib/hwasan/hwasan_report.cc
@@ -16,6 +16,7 @@
 #include "hwasan_allocator.h"
 #include "hwasan_mapping.h"
 #include "hwasan_thread.h"
+#include "hwasan_thread_list.h"
 #include "sanitizer_common/sanitizer_allocator_internal.h"
 #include "sanitizer_common/sanitizer_common.h"
 #include "sanitizer_common/sanitizer_flags.h"
@@ -35,6 +36,31 @@
   return res;
 }
 
+// A RAII object that holds a copy of the current thread stack ring buffer.
+// The actual stack buffer may change while we are iterating over it (for
+// example, Printf may call syslog() which can itself be built with hwasan).
+class SavedStackAllocations {
+ public:
+  SavedStackAllocations(StackAllocationsRingBuffer *rb) {
+    uptr size = rb->size() * sizeof(uptr);
+    void *storage =
+        MmapAlignedOrDieOnFatalError(size, size * 2, "saved stack allocations");
+    new (&rb_) StackAllocationsRingBuffer(*rb, storage);
+  }
+
+  ~SavedStackAllocations() {
+    StackAllocationsRingBuffer *rb = get();
+    UnmapOrDie(rb->StartOfStorage(), rb->size() * sizeof(uptr));
+  }
+
+  StackAllocationsRingBuffer *get() {
+    return (StackAllocationsRingBuffer *)&rb_;
+  }
+
+ private:
+  uptr rb_;
+};
+
 class Decorator: public __sanitizer::SanitizerCommonDecorator {
  public:
   Decorator() : SanitizerCommonDecorator() { }
@@ -63,7 +89,9 @@
   return 0;
 }
 
-void PrintAddressDescription(uptr tagged_addr, uptr access_size) {
+void PrintAddressDescription(
+    uptr tagged_addr, uptr access_size,
+    StackAllocationsRingBuffer *current_stack_allocations) {
   Decorator d;
   int num_descriptions_printed = 0;
   uptr untagged_addr = UntagAddr(tagged_addr);
@@ -109,7 +137,7 @@
     }
   }
 
-  Thread::VisitAllLiveThreads([&](Thread *t) {
+  hwasanThreadList().VisitAllLiveThreads([&](Thread *t) {
     // Scan all threads' ring buffers to find if it's a heap-use-after-free.
     HeapAllocationRecord har;
     if (uptr D = FindHeapAllocation(t->heap_allocations(), tagged_addr, &har)) {
@@ -145,6 +173,25 @@
       Printf("%s", d.Default());
       t->Announce();
 
+      // Temporary report section, needs to be improved.
+      Printf("Previosly allocated frames:\n");
+      auto *sa = (t == GetCurrentThread() && current_stack_allocations)
+                     ? current_stack_allocations
+                     : t->stack_allocations();
+      uptr frames = Min((uptr)flags()->stack_history_size, sa->size());
+      for (uptr i = 0; i < frames; i++) {
+        uptr record = (*sa)[i];
+        if (!record)
+          break;
+        uptr sp = (record >> 48) << 4;
+        uptr pc_mask = (1ULL << 48) - 1;
+        uptr pc = record & pc_mask;
+        uptr fixed_pc = StackTrace::GetNextInstructionPc(pc);
+        StackTrace stack(&fixed_pc, 1);
+        Printf("record: %p pc: %p sp: %p", record, pc, sp);
+        stack.Print();
+      }
+
       num_descriptions_printed++;
     }
   });
@@ -170,13 +217,16 @@
 void ReportInvalidAccessInsideAddressRange(const char *what, const void *start,
                                            uptr size, uptr offset) {
   ScopedErrorReportLock l;
+  SavedStackAllocations current_stack_allocations(
+      GetCurrentThread()->stack_allocations());
 
   Decorator d;
   Printf("%s", d.Warning());
   Printf("%sTag mismatch in %s%s%s at offset %zu inside [%p, %zu)%s\n",
          d.Warning(), d.Name(), what, d.Warning(), offset, start, size,
          d.Default());
-  PrintAddressDescription((uptr)start + offset, 1);
+  PrintAddressDescription((uptr)start + offset, 1,
+                          current_stack_allocations.get());
   // if (__sanitizer::Verbosity())
   //   DescribeMemoryRange(start, size);
 }
@@ -224,7 +274,7 @@
 
   stack->Print();
 
-  PrintAddressDescription(tagged_addr, 0);
+  PrintAddressDescription(tagged_addr, 0, nullptr);
 
   PrintTagsAroundAddr(tag_ptr);
 
@@ -235,6 +285,8 @@
 void ReportTagMismatch(StackTrace *stack, uptr tagged_addr, uptr access_size,
                        bool is_store) {
   ScopedErrorReportLock l;
+  SavedStackAllocations current_stack_allocations(
+      GetCurrentThread()->stack_allocations());
 
   Decorator d;
   Printf("%s", d.Error());
@@ -258,7 +310,8 @@
 
   stack->Print();
 
-  PrintAddressDescription(tagged_addr, access_size);
+  PrintAddressDescription(tagged_addr, access_size,
+                          current_stack_allocations.get());
   t->Announce();
 
   PrintTagsAroundAddr(tag_ptr);
Index: compiler-rt/trunk/lib/hwasan/hwasan_thread.h
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_thread.h
+++ compiler-rt/trunk/lib/hwasan/hwasan_thread.h
@@ -16,12 +16,15 @@
 
 #include "hwasan_allocator.h"
 #include "sanitizer_common/sanitizer_common.h"
+#include "sanitizer_common/sanitizer_ring_buffer.h"
 
 namespace __hwasan {
 
+typedef __sanitizer::CompactRingBuffer<uptr> StackAllocationsRingBuffer;
+
 class Thread {
  public:
-  static void Create();  // Must be called from the thread itself.
+  void Init(uptr stack_buffer_start, uptr stack_buffer_size);  // Must be called from the thread itself.
   void Destroy();
 
   uptr stack_top() { return stack_top_; }
@@ -48,27 +51,15 @@
   void LeaveInterceptorScope() { in_interceptor_scope_--; }
 
   AllocatorCache *allocator_cache() { return &allocator_cache_; }
-  HeapAllocationsRingBuffer *heap_allocations() {
-    return heap_allocations_;
-  }
+  HeapAllocationsRingBuffer *heap_allocations() { return heap_allocations_; }
+  StackAllocationsRingBuffer *stack_allocations() { return stack_allocations_; }
 
   tag_t GenerateRandomTag();
 
-  int destructor_iterations_;
   void DisableTagging() { tagging_disabled_++; }
   void EnableTagging() { tagging_disabled_--; }
   bool TaggingIsDisabled() const { return tagging_disabled_; }
 
-  template <class CB>
-  static void VisitAllLiveThreads(CB cb) {
-    SpinMutexLock l(&thread_list_mutex);
-    Thread *t = thread_list_head;
-    while (t) {
-      cb(t);
-      t = t->next_;
-    }
-  }
-
   u64 unique_id() const { return unique_id_; }
   void Announce() {
     if (announced_) return;
@@ -76,22 +67,9 @@
     Print("Thread: ");
   }
 
-  struct ThreadStats {
-    uptr n_live_threads;
-    uptr total_stack_size;
-  };
-
-  static ThreadStats GetThreadStats() {
-    SpinMutexLock l(&thread_list_mutex);
-    return thread_stats;
-  }
-
-  static uptr MemoryUsedPerThread();
-
  private:
   // NOTE: There is no Thread constructor. It is allocated
   // via mmap() and *must* be valid in zero-initialized state.
-  void Init();
   void ClearShadowForThreadStackAndTLS();
   void Print(const char *prefix);
   uptr stack_top_;
@@ -108,23 +86,23 @@
 
   AllocatorCache allocator_cache_;
   HeapAllocationsRingBuffer *heap_allocations_;
+  StackAllocationsRingBuffer *stack_allocations_;
 
   static void InsertIntoThreadList(Thread *t);
   static void RemoveFromThreadList(Thread *t);
   Thread *next_;  // All live threads form a linked list.
-  static SpinMutex thread_list_mutex;
-  static Thread *thread_list_head;
-  static ThreadStats thread_stats;
 
   u64 unique_id_;  // counting from zero.
 
   u32 tagging_disabled_;  // if non-zero, malloc uses zero tag in this thread.
 
   bool announced_;
+
+  friend struct ThreadListHead;
 };
 
 Thread *GetCurrentThread();
-void SetCurrentThread(Thread *t);
+uptr *GetCurrentThreadLongPtr();
 
 struct ScopedTaggingDisabler {
   ScopedTaggingDisabler() { GetCurrentThread()->DisableTagging(); }
Index: compiler-rt/trunk/lib/hwasan/hwasan_thread.cc
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_thread.cc
+++ compiler-rt/trunk/lib/hwasan/hwasan_thread.cc
@@ -9,6 +9,7 @@
 #include "sanitizer_common/sanitizer_placement_new.h"
 #include "sanitizer_common/sanitizer_tls_get_addr.h"
 
+
 namespace __hwasan {
 
 static u32 RandomSeed() {
@@ -24,102 +25,52 @@
   return seed;
 }
 
-Thread *Thread::thread_list_head;
-SpinMutex Thread::thread_list_mutex;
-Thread::ThreadStats Thread::thread_stats;
-
-void Thread::InsertIntoThreadList(Thread *t) {
-  CHECK(!t->next_);
-  SpinMutexLock l(&thread_list_mutex);
-  thread_stats.n_live_threads++;
-  thread_stats.total_stack_size += t->stack_size();
-  if (!thread_list_head) {
-    thread_list_head = t;
-    return;
-  }
-  Thread *last = thread_list_head;
-  while (last->next_)
-    last = last->next_;
-  last->next_ = t;
-}
-
-void Thread::RemoveFromThreadList(Thread *t) {
-  SpinMutexLock l(&thread_list_mutex);
-  thread_stats.n_live_threads--;
-  thread_stats.total_stack_size -= t->stack_size();
-  if (t == thread_list_head) {
-    thread_list_head = t->next_;
-    t->next_ = nullptr;
-    return;
-  }
-  Thread *prev = thread_list_head;
-  Thread *cur = prev->next_;
-  CHECK(cur);
-  while (cur) {
-    if (cur == t) {
-      prev->next_ = cur->next_;
-      return;
-    }
-    prev = cur;
-    cur = cur->next_;
-  }
-  CHECK(0 && "RemoveFromThreadList: thread not found");
-}
-
-void Thread::Create() {
+void Thread::Init(uptr stack_buffer_start, uptr stack_buffer_size) {
   static u64 unique_id;
-  uptr PageSize = GetPageSizeCached();
-  uptr size = RoundUpTo(sizeof(Thread), PageSize);
-  Thread *thread = (Thread*)MmapOrDie(size, __func__);
-  thread->destructor_iterations_ = GetPthreadDestructorIterations();
-  thread->unique_id_ = unique_id++;
-  thread->random_state_ =
-      flags()->random_tags ? RandomSeed() : thread->unique_id_;
+  unique_id_ = unique_id++;
+  random_state_ = flags()->random_tags ? RandomSeed() : unique_id_;
   if (auto sz = flags()->heap_history_size)
-    thread->heap_allocations_ = HeapAllocationsRingBuffer::New(sz);
-  SetCurrentThread(thread);
-  thread->Init();
-  InsertIntoThreadList(thread);
-}
+    heap_allocations_ = HeapAllocationsRingBuffer::New(sz);
 
-uptr Thread::MemoryUsedPerThread() {
-  uptr res = sizeof(Thread);
-  if (auto sz = flags()->heap_history_size)
-    res += HeapAllocationsRingBuffer::SizeInBytes(sz);
-  return res;
-}
+  HwasanTSDThreadInit();  // Only needed with interceptors.
+  uptr *ThreadLong = GetCurrentThreadLongPtr();
+  // The following implicitly sets (this) as the current thread.
+  stack_allocations_ = new (ThreadLong)
+      StackAllocationsRingBuffer((void *)stack_buffer_start, stack_buffer_size);
+  // Check that it worked.
+  CHECK_EQ(GetCurrentThread(), this);
 
-void Thread::Init() {
-  // GetPthreadDestructorIterations may call malloc, so disable the tagging.
+  // ScopedTaggingDisable needs GetCurrentThread to be set up.
   ScopedTaggingDisabler disabler;
 
   // If this process is "init" (pid 1), /proc may not be mounted yet.
   if (IsMainThread() && !FileExists("/proc/self/maps")) {
     stack_top_ = stack_bottom_ = 0;
     tls_begin_ = tls_end_ = 0;
-    return;
-  }
+  } else {
+    uptr tls_size;
+    uptr stack_size;
+    GetThreadStackAndTls(IsMainThread(), &stack_bottom_, &stack_size,
+                         &tls_begin_, &tls_size);
+    stack_top_ = stack_bottom_ + stack_size;
+    tls_end_ = tls_begin_ + tls_size;
 
-  uptr tls_size;
-  uptr stack_size;
-  GetThreadStackAndTls(IsMainThread(), &stack_bottom_, &stack_size, &tls_begin_,
-                       &tls_size);
-  stack_top_ = stack_bottom_ + stack_size;
-  tls_end_ = tls_begin_ + tls_size;
-
-  int local;
-  CHECK(AddrIsInStack((uptr)&local));
-  CHECK(MemIsApp(stack_bottom_));
-  CHECK(MemIsApp(stack_top_ - 1));
-
-  if (stack_bottom_) {
+    int local;
+    CHECK(AddrIsInStack((uptr)&local));
     CHECK(MemIsApp(stack_bottom_));
     CHECK(MemIsApp(stack_top_ - 1));
+
+    if (stack_bottom_) {
+      CHECK(MemIsApp(stack_bottom_));
+      CHECK(MemIsApp(stack_top_ - 1));
+    }
   }
+
   if (flags()->verbose_threads) {
     if (IsMainThread()) {
-      Printf("sizeof(Thread): %zd sizeof(RB): %zd\n", sizeof(Thread),
-             heap_allocations_->SizeInBytes());
+      Printf("sizeof(Thread): %zd sizeof(HeapRB): %zd sizeof(StackRB): %zd\n",
+             sizeof(Thread), heap_allocations_->SizeInBytes(),
+             stack_allocations_->size() * sizeof(uptr));
     }
     Print("Creating  : ");
   }
@@ -137,11 +88,8 @@
     Print("Destroying: ");
   AllocatorSwallowThreadLocalCache(allocator_cache());
   ClearShadowForThreadStackAndTLS();
-  RemoveFromThreadList(this);
-  uptr size = RoundUpTo(sizeof(Thread), GetPageSizeCached());
   if (heap_allocations_)
     heap_allocations_->Delete();
-  UnmapOrDie(this, size);
   DTLS_Destroy();
 }
 
Index: compiler-rt/trunk/lib/hwasan/hwasan_thread_list.h
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_thread_list.h
+++ compiler-rt/trunk/lib/hwasan/hwasan_thread_list.h
@@ -0,0 +1,200 @@
+//===-- hwasan_thread_list.h ------------------------------------*- 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 HWAddressSanitizer.
+//
+//===----------------------------------------------------------------------===//
+
+// HwasanThreadList is a registry for live threads, as well as an allocator for
+// HwasanThread objects and their stack history ring buffers. There are
+// constraints on memory layout of the shadow region and CompactRingBuffer that
+// are part of the ABI contract between compiler-rt and llvm.
+//
+// * Start of the shadow memory region is aligned to 2**kShadowBaseAlignment.
+// * All stack ring buffers are located within (2**kShadowBaseAlignment)
+// sized region below and adjacent to the shadow region.
+// * Each ring buffer has a size of (2**N)*4096 where N is in [0, 8), and is
+// aligned to twice its size. The value of N can be different for each buffer.
+//
+// These constrains guarantee that, given an address A of any element of the
+// ring buffer,
+//     A_next = (A + sizeof(uptr)) & ~((1 << (N + 13)) - 1)
+//   is the address of the next element of that ring buffer (with wrap-around).
+// And, with K = kShadowBaseAlignment,
+//     S = (A | ((1 << K) - 1)) + 1
+//   (align up to kShadowBaseAlignment) is the start of the shadow region.
+//
+// These calculations are used in compiler instrumentation to update the ring
+// buffer and obtain the base address of shadow using only two inputs: address
+// of the current element of the ring buffer, and N (i.e. size of the ring
+// buffer). Since the value of N is very limited, we pack both inputs into a
+// single thread-local word as
+//   (1 << (N + 56)) | A
+// See the implementation of class CompactRingBuffer, which is what is stored in
+// said thread-local word.
+//
+// Note the unusual way of aligning up the address of the shadow:
+//   (A | ((1 << K) - 1)) + 1
+// It is only correct if A is not already equal to the shadow base address, but
+// it saves 2 instructions on AArch64.
+
+#include "hwasan.h"
+#include "hwasan_allocator.h"
+#include "hwasan_flags.h"
+#include "hwasan_thread.h"
+
+#include "sanitizer_common/sanitizer_placement_new.h"
+
+namespace __hwasan {
+
+static uptr RingBufferSize() {
+  uptr desired_bytes = flags()->stack_history_size * sizeof(uptr);
+  // FIXME: increase the limit to 8 once this bug is fixed:
+  // https://bugs.llvm.org/show_bug.cgi?id=39030
+  for (int shift = 1; shift < 7; ++shift) {
+    uptr size = 4096 * (1ULL << shift);
+    if (size >= desired_bytes)
+      return size;
+  }
+  Printf("stack history size too large: %d\n", flags()->stack_history_size);
+  CHECK(0);
+  return 0;
+}
+
+struct ThreadListHead {
+  Thread *list_;
+
+  ThreadListHead() : list_(nullptr) {}
+
+  void Push(Thread *t) {
+    t->next_ = list_;
+    list_ = t;
+  }
+
+  Thread *Pop() {
+    Thread *t = list_;
+    if (t)
+      list_ = t->next_;
+    return t;
+  }
+
+  void Remove(Thread *t) {
+    Thread **cur = &list_;
+    while (*cur != t) cur = &(*cur)->next_;
+    CHECK(*cur && "thread not found");
+    *cur = (*cur)->next_;
+  }
+
+  template <class CB>
+  void ForEach(CB cb) {
+    Thread *t = list_;
+    while (t) {
+      cb(t);
+      t = t->next_;
+    }
+  }
+};
+
+struct ThreadStats {
+  uptr n_live_threads;
+  uptr total_stack_size;
+};
+
+class HwasanThreadList {
+ public:
+  HwasanThreadList(uptr storage, uptr size)
+      : free_space_(storage),
+        free_space_end_(storage + size),
+        ring_buffer_size_(RingBufferSize()) {}
+
+  Thread *CreateCurrentThread() {
+    Thread *t;
+    {
+      SpinMutexLock l(&list_mutex_);
+      t = free_list_.Pop();
+      if (t)
+        internal_memset((void *)t, 0, sizeof(Thread) + ring_buffer_size_);
+      else
+        t = AllocThread();
+      live_list_.Push(t);
+    }
+    t->Init((uptr)(t + 1), ring_buffer_size_);
+    AddThreadStats(t);
+    return t;
+  }
+
+  void ReleaseThread(Thread *t) {
+    // FIXME: madvise away the ring buffer?
+    RemoveThreadStats(t);
+    t->Destroy();
+    SpinMutexLock l(&list_mutex_);
+    live_list_.Remove(t);
+    free_list_.Push(t);
+  }
+
+  Thread *GetThreadByBufferAddress(uptr p) {
+    uptr align = ring_buffer_size_ * 2;
+    return (Thread *)(RoundDownTo(p, align) - sizeof(Thread));
+  }
+
+  uptr MemoryUsedPerThread() {
+    uptr res = sizeof(Thread) + ring_buffer_size_;
+    if (auto sz = flags()->heap_history_size)
+      res += HeapAllocationsRingBuffer::SizeInBytes(sz);
+    return res;
+  }
+
+  template <class CB>
+  void VisitAllLiveThreads(CB cb) {
+    SpinMutexLock l(&list_mutex_);
+    live_list_.ForEach(cb);
+  }
+
+  void AddThreadStats(Thread *t) {
+    SpinMutexLock l(&stats_mutex_);
+    stats_.n_live_threads++;
+    stats_.total_stack_size += t->stack_size();
+  }
+
+  void RemoveThreadStats(Thread *t) {
+    SpinMutexLock l(&stats_mutex_);
+    stats_.n_live_threads--;
+    stats_.total_stack_size -= t->stack_size();
+  }
+
+  ThreadStats GetThreadStats() {
+    SpinMutexLock l(&stats_mutex_);
+    return stats_;
+  }
+
+ private:
+  Thread *AllocThread() {
+    uptr align = ring_buffer_size_ * 2;
+    uptr ring_buffer_start = RoundUpTo(free_space_ + sizeof(Thread), align);
+    free_space_ = ring_buffer_start + ring_buffer_size_;
+    CHECK(free_space_ <= free_space_end_ && "out of thread memory");
+    return (Thread *)(ring_buffer_start - sizeof(Thread));
+  }
+
+  uptr free_space_;
+  uptr free_space_end_;
+  uptr ring_buffer_size_;
+
+  ThreadListHead free_list_;
+  ThreadListHead live_list_;
+  SpinMutex list_mutex_;
+
+  ThreadStats stats_;
+  SpinMutex stats_mutex_;
+};
+
+void InitThreadList(uptr storage, uptr size);
+HwasanThreadList &hwasanThreadList();
+
+} // namespace
Index: compiler-rt/trunk/lib/hwasan/hwasan_thread_list.cc
===================================================================
--- compiler-rt/trunk/lib/hwasan/hwasan_thread_list.cc
+++ compiler-rt/trunk/lib/hwasan/hwasan_thread_list.cc
@@ -0,0 +1,15 @@
+#include "hwasan_thread_list.h"
+
+namespace __hwasan {
+static ALIGNED(16) char thread_list_placeholder[sizeof(HwasanThreadList)];
+static HwasanThreadList *hwasan_thread_list;
+
+HwasanThreadList &hwasanThreadList() { return *hwasan_thread_list; }
+
+void InitThreadList(uptr storage, uptr size) {
+  CHECK(hwasan_thread_list == nullptr);
+  hwasan_thread_list =
+      new (thread_list_placeholder) HwasanThreadList(storage, size);
+}
+
+} // namespace
Index: compiler-rt/trunk/lib/sanitizer_common/sanitizer_ring_buffer.h
===================================================================
--- compiler-rt/trunk/lib/sanitizer_common/sanitizer_ring_buffer.h
+++ compiler-rt/trunk/lib/sanitizer_common/sanitizer_ring_buffer.h
@@ -72,12 +72,91 @@
   // L: last_, always points to the last data element.
   // N: next_, initially equals to last_, is decremented on every push,
   //    wraps around if it's less or equal than its own address.
-
   T *last_;
   T *next_;
   T data_[1];  // flexible array.
 };
 
+// A ring buffer with externally provided storage that encodes its state in 8
+// bytes. Has significant constraints on size and alignment of storage.
+// See a comment in hwasan/hwasan_thread_list.h for the motivation behind this.
+#if SANITIZER_WORDSIZE == 64
+template <class T>
+class CompactRingBuffer {
+  // Top byte of long_ stores the buffer size in pages.
+  // Lower bytes store the address of the next buffer element.
+  static constexpr int kPageSizeBits = 12;
+  static constexpr int kSizeShift = 56;
+  static constexpr uptr kNextMask = (1ULL << kSizeShift) - 1;
+
+  uptr GetStorageSize() const { return (long_ >> kSizeShift) << kPageSizeBits; }
+
+  void Init(void *storage, uptr size) {
+    CHECK_EQ(sizeof(CompactRingBuffer<T>), sizeof(void *));
+    CHECK(IsPowerOfTwo(size));
+    CHECK_GE(size, 1 << kPageSizeBits);
+    CHECK_LE(size, 128 << kPageSizeBits);
+    CHECK_EQ(size % 4096, 0);
+    CHECK_EQ(size % sizeof(T), 0);
+    CHECK_EQ((uptr)storage % (size * 2), 0);
+    long_ = (uptr)storage | ((size >> kPageSizeBits) << kSizeShift);
+  }
+
+  void SetNext(const T *next) {
+    long_ = (long_ & ~kNextMask) | (uptr)next;
+  }
+
+ public:
+  CompactRingBuffer(void *storage, uptr size) {
+    Init(storage, size);
+  }
+
+  // A copy constructor of sorts.
+  CompactRingBuffer(const CompactRingBuffer &other, void *storage) {
+    uptr size = other.GetStorageSize();
+    internal_memcpy(storage, other.StartOfStorage(), size);
+    Init(storage, size);
+    uptr Idx = other.Next() - (const T *)other.StartOfStorage();
+    SetNext((const T *)storage + Idx);
+  }
+
+  T *Next() const { return (T *)(long_ & kNextMask); }
+
+  void *StartOfStorage() const {
+    return (void *)((uptr)Next() & ~(GetStorageSize() - 1));
+  }
+
+  void *EndOfStorage() const {
+    return (void *)((uptr)StartOfStorage() + GetStorageSize());
+  }
+
+  uptr size() const { return GetStorageSize() / sizeof(T); }
+
+  void push(T t) {
+    T *next = Next();
+    *next = t;
+    next++;
+    next = (T *)((uptr)next & ~GetStorageSize());
+    SetNext(next);
+  }
+
+  T operator[](uptr Idx) const {
+    CHECK_LT(Idx, size());
+    const T *Begin = (const T *)StartOfStorage();
+    sptr StorageIdx = Next() - Begin;
+    StorageIdx -= (sptr)(Idx + 1);
+    if (StorageIdx < 0)
+      StorageIdx += size();
+    return Begin[StorageIdx];
+  }
+
+ public:
+  ~CompactRingBuffer() {}
+  CompactRingBuffer(const CompactRingBuffer &) = delete;
+
+  uptr long_;
+};
+#endif
 }  // namespace __sanitizer
 
 #endif  // SANITIZER_RING_BUFFER_H
Index: compiler-rt/trunk/lib/sanitizer_common/tests/sanitizer_ring_buffer_test.cc
===================================================================
--- compiler-rt/trunk/lib/sanitizer_common/tests/sanitizer_ring_buffer_test.cc
+++ compiler-rt/trunk/lib/sanitizer_common/tests/sanitizer_ring_buffer_test.cc
@@ -66,6 +66,7 @@
 #undef EXPECT_RING_BUFFER
 }
 
+#if SANITIZER_WORDSIZE == 64
 TEST(RingBuffer, int64) {
   TestRB<int64_t>();
 }
@@ -74,4 +75,25 @@
   TestRB<LargeStruct>();
 }
 
+template<typename T>
+CompactRingBuffer<T> *AllocCompactRingBuffer(size_t count) {
+  size_t sz = sizeof(T) * count;
+  EXPECT_EQ(0ULL, sz % 4096);
+  void *p = MmapAlignedOrDieOnFatalError(sz, sz * 2, "CompactRingBuffer");
+  return new CompactRingBuffer<T>(p, sz);
+}
+
+TEST(CompactRingBuffer, int64) {
+  const size_t page_sizes[] = {1, 2, 4, 128};
+
+  for (size_t pages : page_sizes) {
+    size_t count = 4096 * pages / sizeof(int64_t);
+    auto R = AllocCompactRingBuffer<int64_t>(count);
+    int64_t top = count * 3 + 13;
+    for (int64_t i = 0; i < top; ++i) R->push(i);
+    for (int64_t i = 0; i < (int64_t)count; ++i)
+      EXPECT_EQ(top - i - 1, (*R)[i]);
+  }
+}
+#endif
 }  // namespace __sanitizer
Index: compiler-rt/trunk/test/hwasan/TestCases/deep-recursion.c
===================================================================
--- compiler-rt/trunk/test/hwasan/TestCases/deep-recursion.c
+++ compiler-rt/trunk/test/hwasan/TestCases/deep-recursion.c
@@ -0,0 +1,73 @@
+// RUN: %clang_hwasan -O1 %s -o %t
+// RUN: %env_hwasan_opts=stack_history_size=1 not %run %t 2>&1 | FileCheck %s --check-prefix=D1
+// RUN: %env_hwasan_opts=stack_history_size=2 not %run %t 2>&1 | FileCheck %s --check-prefix=D2
+// RUN: %env_hwasan_opts=stack_history_size=3 not %run %t 2>&1 | FileCheck %s --check-prefix=D3
+// RUN: %env_hwasan_opts=stack_history_size=5 not %run %t 2>&1 | FileCheck %s --check-prefix=D5
+// RUN:                                       not %run %t 2>&1 | FileCheck %s --check-prefix=DEFAULT
+
+// REQUIRES: stable-runtime
+
+#include <stdlib.h>
+// At least -O1 is needed for this function to not have a stack frame on
+// AArch64.
+void USE(void *x) { // pretend_to_do_something(void *x)
+  __asm__ __volatile__("" : : "r" (x) : "memory");
+}
+
+volatile int four = 4;
+
+__attribute__((noinline)) void OOB() { int x[4]; x[four] = 0; USE(&x[0]); }
+__attribute__((noinline)) void FUNC1() { int x; USE(&x); OOB(); }
+__attribute__((noinline)) void FUNC2() { int x; USE(&x); FUNC1(); }
+__attribute__((noinline)) void FUNC3() { int x; USE(&x); FUNC2(); }
+__attribute__((noinline)) void FUNC4() { int x; USE(&x); FUNC3(); }
+__attribute__((noinline)) void FUNC5() { int x; USE(&x); FUNC4(); }
+__attribute__((noinline)) void FUNC6() { int x; USE(&x); FUNC5(); }
+__attribute__((noinline)) void FUNC7() { int x; USE(&x); FUNC6(); }
+__attribute__((noinline)) void FUNC8() { int x; USE(&x); FUNC7(); }
+__attribute__((noinline)) void FUNC9() { int x; USE(&x); FUNC8(); }
+__attribute__((noinline)) void FUNC10() { int x; USE(&x); FUNC9(); }
+
+int main() { FUNC10(); }
+
+// D1: Previosly allocated frames
+// D1: in OOB
+// D1-NOT: in FUNC
+// D1: Memory tags around the buggy address
+
+// D2: Previosly allocated frames
+// D2: in OOB
+// D2: in FUNC1
+// D2-NOT: in FUNC
+// D2: Memory tags around the buggy address
+
+// D3: Previosly allocated frames
+// D3: in OOB
+// D3: in FUNC1
+// D3: in FUNC2
+// D3-NOT: in FUNC
+// D3: Memory tags around the buggy address
+
+// D5: Previosly allocated frames
+// D5: in OOB
+// D5: in FUNC1
+// D5: in FUNC2
+// D5: in FUNC3
+// D5: in FUNC4
+// D5-NOT: in FUNC
+// D5: Memory tags around the buggy address
+
+// DEFAULT: Previosly allocated frames
+// DEFAULT: in OOB
+// DEFAULT: in FUNC1
+// DEFAULT: in FUNC2
+// DEFAULT: in FUNC3
+// DEFAULT: in FUNC4
+// DEFAULT: in FUNC5
+// DEFAULT: in FUNC6
+// DEFAULT: in FUNC7
+// DEFAULT: in FUNC8
+// DEFAULT: in FUNC9
+// DEFAULT: in FUNC10
+// DEFAULT-NOT: in FUNC
+// DEFAULT: Memory tags around the buggy address
Index: compiler-rt/trunk/test/hwasan/TestCases/rich-stack.c
===================================================================
--- compiler-rt/trunk/test/hwasan/TestCases/rich-stack.c
+++ compiler-rt/trunk/test/hwasan/TestCases/rich-stack.c
@@ -0,0 +1,66 @@
+// Test how stack frames are reported (not fully implemented yet).
+// RUN: %clang_hwasan %s -o %t
+// RUN: not %run %t 3 2 -1 2>&1 | FileCheck %s --check-prefix=R321
+// REQUIRES: stable-runtime
+#include <stdint.h>
+#include <stdlib.h>
+void USE(void *x) { // pretend_to_do_something(void *x)
+  __asm__ __volatile__("" : : "r" (x) : "memory");
+}
+void USE2(void *a, void *b) { USE(a); USE(b); }
+void USE4(void *a, void *b, void *c, void *d) { USE2(a, b); USE2(c, d); }
+
+void BAR(int depth, int err_depth, int offset);
+
+uint64_t *leaked_ptr;
+
+void FOO(int depth, int err_depth, int offset) {
+  uint8_t v1;
+  uint16_t v2;
+  uint32_t v4;
+  uint64_t v8;
+  uint64_t v16[2];
+  uint64_t v32[4];
+  uint64_t v48[3];
+  USE4(&v1, &v2, &v4, &v8);  USE4(&v16, &v32, &v48, 0);
+  leaked_ptr = &v16[0];
+  if (depth)
+    BAR(depth - 1, err_depth, offset);
+
+  if (err_depth == depth)
+    v16[offset] = 0;  // maybe OOB.
+  if (err_depth == -depth)
+    leaked_ptr[offset] = 0; // maybe UAR.
+  USE(&v16);
+}
+
+void BAR(int depth, int err_depth, int offset) {
+  uint64_t x16[2];
+  uint64_t x32[4];
+  USE2(&x16, &x32);
+  leaked_ptr = &x16[0];
+  if (depth)
+    FOO(depth - 1, err_depth, offset);
+  if (err_depth == depth)
+    x16[offset] = 0;  // maybe OOB
+  if (err_depth == -depth)
+    leaked_ptr[offset] = 0;  // maybe UAR
+  USE(&x16);
+}
+
+
+int main(int argc, char **argv) {
+  if (argc != 4) return -1;
+  int depth = atoi(argv[1]);
+  int err_depth = atoi(argv[2]);
+  int offset = atoi(argv[3]);
+  FOO(depth, err_depth, offset);
+  return 0;
+}
+
+// R321: HWAddressSanitizer: tag-mismatch
+// R321-NEXT: WRITE of size 8
+// R321-NEXT: in BAR
+// R321-NEXT: in FOO
+// R321-NEXT: in main
+// R321: is located in stack of thread T0
Index: compiler-rt/trunk/test/hwasan/TestCases/stack-history-length.c
===================================================================
--- compiler-rt/trunk/test/hwasan/TestCases/stack-history-length.c
+++ compiler-rt/trunk/test/hwasan/TestCases/stack-history-length.c
@@ -0,0 +1,36 @@
+// RUN: %clang_hwasan -O1 -DX=2046 %s -o %t.2046
+// RUN: %clang_hwasan -O1 -DX=2047 %s -o %t.2047
+// RUN: %env_hwasan_opts=stack_history_size=2048 not %run %t.2046 2>&1 | FileCheck %s --check-prefix=YES
+// RUN: %env_hwasan_opts=stack_history_size=2048 not %run %t.2047 2>&1 | FileCheck %s --check-prefix=NO
+
+// REQUIRES: stable-runtime
+
+#include <stdlib.h>
+
+void USE(void *x) { // pretend_to_do_something(void *x)
+  __asm__ __volatile__("" : : "r" (x) : "memory");
+}
+
+volatile int four = 4;
+__attribute__((noinline)) void FUNC0() { int x[4]; USE(&x[0]); }
+__attribute__((noinline)) void FUNC() { int x[4]; USE(&x[0]); }
+__attribute__((noinline)) void OOB() { int x[4]; x[four] = 0; USE(&x[0]); }
+
+int main() {
+  // FUNC0 is X+2's element of the ring buffer.
+  // If runtime buffer size is less than it, FUNC0 record will be lost.
+  FUNC0();
+  for (int i = 0; i < X; ++i)
+    FUNC();
+  OOB();
+}
+
+// YES: Previosly allocated frames
+// YES: OOB
+// YES: FUNC
+// YES: FUNC0
+
+// NO: Previosly allocated frames
+// NO: OOB
+// NO: FUNC
+// NO-NOT: FUNC0
Index: llvm/trunk/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp
===================================================================
--- llvm/trunk/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp
+++ llvm/trunk/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp
@@ -63,6 +63,8 @@
     std::numeric_limits<uint64_t>::max();
 static const unsigned kPointerTagShift = 56;
 
+static const unsigned kShadowBaseAlignment = 32;
+
 static cl::opt<std::string> ClMemoryAccessCallbackPrefix(
     "hwasan-memory-access-callback-prefix",
     cl::desc("Prefix for memory access callbacks"), cl::Hidden,
@@ -132,6 +134,18 @@
                 cl::desc("Access dynamic shadow through an ifunc global on "
                          "platforms that support this"),
                 cl::Hidden, cl::init(false));
+
+static cl::opt<bool> ClWithTls(
+    "hwasan-with-tls",
+    cl::desc("Access dynamic shadow through an thread-local pointer on "
+             "platforms that support this"),
+    cl::Hidden, cl::init(true));
+
+static cl::opt<bool>
+    ClRecordStackHistory("hwasan-record-stack-history",
+                         cl::desc("Record stack frames with tagged allocations "
+                                  "in a thread-local ring buffer"),
+                         cl::Hidden, cl::init(true));
 namespace {
 
 /// An instrumentation pass implementing detection of addressability bugs
@@ -155,7 +169,7 @@
 
   void initializeCallbacks(Module &M);
 
-  void maybeInsertDynamicShadowAtFunctionEntry(Function &F);
+  Value *getDynamicShadowNonTls(IRBuilder<> &IRB);
 
   void untagPointerOperand(Instruction *I, Value *Addr);
   Value *memToShadow(Value *Shadow, Type *Ty, IRBuilder<> &IRB);
@@ -172,13 +186,16 @@
   Value *tagPointer(IRBuilder<> &IRB, Type *Ty, Value *PtrLong, Value *Tag);
   Value *untagPointer(IRBuilder<> &IRB, Value *PtrLong);
   bool instrumentStack(SmallVectorImpl<AllocaInst *> &Allocas,
-                       SmallVectorImpl<Instruction *> &RetVec);
+                       SmallVectorImpl<Instruction *> &RetVec, Value *StackTag);
   Value *getNextTagWithCall(IRBuilder<> &IRB);
   Value *getStackBaseTag(IRBuilder<> &IRB);
   Value *getAllocaTag(IRBuilder<> &IRB, Value *StackTag, AllocaInst *AI,
                      unsigned AllocaNo);
   Value *getUARTag(IRBuilder<> &IRB, Value *StackTag);
 
+  Value *getHwasanThreadSlotPtr(IRBuilder<> &IRB, Type *Ty);
+  Value *emitPrologue(IRBuilder<> &IRB, bool WithFrameRecord);
+
 private:
   LLVMContext *C;
   Triple TargetTriple;
@@ -188,10 +205,14 @@
   /// If InGlobal is true, then
   ///   extern char __hwasan_shadow[];
   ///   shadow = (mem >> Scale) + &__hwasan_shadow
+  /// If InTls is true, then
+  ///   extern char *__hwasan_tls;
+  ///   shadow = (mem >> Scale) + align_up(__hwasan_shadow, kShadowBaseAlignment)
   struct ShadowMapping {
     int Scale;
     uint64_t Offset;
     bool InGlobal;
+    bool InTls;
 
     void init(Triple &TargetTriple);
     unsigned getAllocaAlignment() const { return 1U << Scale; }
@@ -216,6 +237,7 @@
   Constant *ShadowGlobal;
 
   Value *LocalDynamicShadow = nullptr;
+  GlobalValue *ThreadPtrGlobal = nullptr;
 };
 
 } // end anonymous namespace
@@ -263,6 +285,12 @@
                                             /*InitArgs=*/{});
     appendToGlobalCtors(M, HwasanCtorFunction, 0);
   }
+  if (!TargetTriple.isAndroid())
+    appendToCompilerUsed(
+        M, ThreadPtrGlobal = new GlobalVariable(
+               M, IntptrTy, false, GlobalVariable::ExternalLinkage, nullptr,
+               "__hwasan_tls", nullptr, GlobalVariable::InitialExecTLSModel));
+
   return true;
 }
 
@@ -297,12 +325,11 @@
                                        ArrayType::get(IRB.getInt8Ty(), 0));
 }
 
-void HWAddressSanitizer::maybeInsertDynamicShadowAtFunctionEntry(Function &F) {
+Value *HWAddressSanitizer::getDynamicShadowNonTls(IRBuilder<> &IRB) {
   // Generate code only when dynamic addressing is needed.
   if (Mapping.Offset != kDynamicShadowSentinel)
-    return;
+    return nullptr;
 
-  IRBuilder<> IRB(&F.front().front());
   if (Mapping.InGlobal) {
     // An empty inline asm with input reg == output reg.
     // An opaque pointer-to-int cast, basically.
@@ -310,11 +337,12 @@
         FunctionType::get(IntptrTy, {ShadowGlobal->getType()}, false),
         StringRef(""), StringRef("=r,0"),
         /*hasSideEffects=*/false);
-    LocalDynamicShadow = IRB.CreateCall(Asm, {ShadowGlobal}, ".hwasan.shadow");
+    return IRB.CreateCall(Asm, {ShadowGlobal}, ".hwasan.shadow");
   } else {
-    Value *GlobalDynamicAddress = F.getParent()->getOrInsertGlobal(
-        kHwasanShadowMemoryDynamicAddress, IntptrTy);
-    LocalDynamicShadow = IRB.CreateLoad(GlobalDynamicAddress);
+    Value *GlobalDynamicAddress =
+        IRB.GetInsertBlock()->getParent()->getParent()->getOrInsertGlobal(
+            kHwasanShadowMemoryDynamicAddress, IntptrTy);
+    return IRB.CreateLoad(GlobalDynamicAddress);
   }
 }
 
@@ -563,7 +591,7 @@
 
 Value *HWAddressSanitizer::getStackBaseTag(IRBuilder<> &IRB) {
   if (ClGenerateTagsWithCalls)
-    return nullptr;
+    return getNextTagWithCall(IRB);
   // FIXME: use addressofreturnaddress (but implement it in aarch64 backend
   // first).
   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
@@ -631,15 +659,88 @@
   return UntaggedPtrLong;
 }
 
+Value *HWAddressSanitizer::getHwasanThreadSlotPtr(IRBuilder<> &IRB, Type *Ty) {
+  Module *M = IRB.GetInsertBlock()->getParent()->getParent();
+  if (TargetTriple.isAArch64() && TargetTriple.isAndroid()) {
+    Function *ThreadPointerFunc =
+        Intrinsic::getDeclaration(M, Intrinsic::thread_pointer);
+    Value *SlotPtr = IRB.CreatePointerCast(
+        IRB.CreateConstGEP1_32(IRB.CreateCall(ThreadPointerFunc), 0x40),
+        Ty->getPointerTo(0));
+    return SlotPtr;
+  }
+  if (ThreadPtrGlobal)
+    return ThreadPtrGlobal;
+
+
+  return nullptr;
+}
+
+Value *HWAddressSanitizer::emitPrologue(IRBuilder<> &IRB,
+                                        bool WithFrameRecord) {
+  if (!Mapping.InTls)
+    return getDynamicShadowNonTls(IRB);
+
+  Value *SlotPtr = getHwasanThreadSlotPtr(IRB, IntptrTy);
+  assert(SlotPtr);
+
+  Value *ThreadLong = IRB.CreateLoad(SlotPtr);
+  // Extract the address field from ThreadLong. Unnecessary on AArch64 with TBI.
+  Value *ThreadLongMaybeUntagged =
+      TargetTriple.isAArch64() ? ThreadLong : untagPointer(IRB, ThreadLong);
+
+  if (WithFrameRecord) {
+    // Prepare ring buffer data.
+    Function *F = IRB.GetInsertBlock()->getParent();
+    auto PC = IRB.CreatePtrToInt(F, IntptrTy);
+    auto GetStackPointerFn =
+        Intrinsic::getDeclaration(F->getParent(), Intrinsic::frameaddress);
+    Value *SP = IRB.CreatePtrToInt(
+        IRB.CreateCall(GetStackPointerFn,
+                       {Constant::getNullValue(IRB.getInt32Ty())}),
+        IntptrTy);
+    // Mix SP and PC. TODO: also add the tag to the mix.
+    // Assumptions:
+    // PC is 0x0000PPPPPPPPPPPP  (48 bits are meaningful, others are zero)
+    // SP is 0xsssssssssssSSSS0  (4 lower bits are zero)
+    // We only really need ~20 lower non-zero bits (SSSS), so we mix like this:
+    //       0xSSSSPPPPPPPPPPPP
+    SP = IRB.CreateShl(SP, 44);
+
+    // Store data to ring buffer.
+    Value *RecordPtr =
+        IRB.CreateIntToPtr(ThreadLongMaybeUntagged, IntptrTy->getPointerTo(0));
+    IRB.CreateStore(IRB.CreateOr(PC, SP), RecordPtr);
+
+    // Update the ring buffer. Top byte of ThreadLong defines the size of the
+    // buffer in pages, it must be a power of two, and the start of the buffer
+    // must be aligned by twice that much. Therefore wrap around of the ring
+    // buffer is simply Addr &= ~((ThreadLong >> 56) << 12).
+    // The use of AShr instead of LShr is due to
+    //   https://bugs.llvm.org/show_bug.cgi?id=39030
+    // Runtime library makes sure not to use the highest bit.
+    Value *WrapMask = IRB.CreateXor(
+        IRB.CreateShl(IRB.CreateAShr(ThreadLong, 56), 12, "", true, true),
+        ConstantInt::get(IntptrTy, (uint64_t)-1));
+    Value *ThreadLongNew = IRB.CreateAnd(
+        IRB.CreateAdd(ThreadLong, ConstantInt::get(IntptrTy, 8)), WrapMask);
+    IRB.CreateStore(ThreadLongNew, SlotPtr);
+  }
+
+  // Get shadow base address by aligning RecordPtr up.
+  // Note: this is not correct if the pointer is already aligned.
+  // Runtime library will make sure this never happens.
+  Value *ShadowBase = IRB.CreateAdd(
+      IRB.CreateOr(
+          ThreadLongMaybeUntagged,
+          ConstantInt::get(IntptrTy, (1UL << kShadowBaseAlignment) - 1)),
+      ConstantInt::get(IntptrTy, 1), "hwasan.shadow");
+  return ShadowBase;
+}
+
 bool HWAddressSanitizer::instrumentStack(
     SmallVectorImpl<AllocaInst *> &Allocas,
-    SmallVectorImpl<Instruction *> &RetVec) {
-  Function *F = Allocas[0]->getParent()->getParent();
-  Instruction *InsertPt = &*F->getEntryBlock().begin();
-  IRBuilder<> IRB(InsertPt);
-
-  Value *StackTag = getStackBaseTag(IRB);
-
+    SmallVectorImpl<Instruction *> &RetVec, Value *StackTag) {
   // Ideally, we want to calculate tagged stack base pointer, and rewrite all
   // alloca addresses using that. Unfortunately, offsets are not known yet
   // (unless we use ASan-style mega-alloca). Instead we keep the base tag in a
@@ -647,7 +748,7 @@
   // This generates one extra instruction per alloca use.
   for (unsigned N = 0; N < Allocas.size(); ++N) {
     auto *AI = Allocas[N];
-    IRB.SetInsertPoint(AI->getNextNode());
+    IRBuilder<> IRB(AI->getNextNode());
 
     // Replace uses of the alloca with tagged address.
     Value *Tag = getAllocaTag(IRB, StackTag, AI, N);
@@ -702,12 +803,6 @@
 
   LLVM_DEBUG(dbgs() << "Function: " << F.getName() << "\n");
 
-  initializeCallbacks(*F.getParent());
-
-  assert(!LocalDynamicShadow);
-  maybeInsertDynamicShadowAtFunctionEntry(F);
-
-  bool Changed = false;
   SmallVector<Instruction*, 16> ToInstrument;
   SmallVector<AllocaInst*, 8> AllocasToInstrument;
   SmallVector<Instruction*, 8> RetVec;
@@ -740,8 +835,25 @@
     }
   }
 
-  if (!AllocasToInstrument.empty())
-    Changed |= instrumentStack(AllocasToInstrument, RetVec);
+  if (AllocasToInstrument.empty() && ToInstrument.empty())
+    return false;
+
+  initializeCallbacks(*F.getParent());
+
+  assert(!LocalDynamicShadow);
+
+  Instruction *InsertPt = &*F.getEntryBlock().begin();
+  IRBuilder<> EntryIRB(InsertPt);
+  LocalDynamicShadow = emitPrologue(EntryIRB,
+                                    /*WithFrameRecord*/ ClRecordStackHistory &&
+                                        !AllocasToInstrument.empty());
+
+  bool Changed = false;
+  if (!AllocasToInstrument.empty()) {
+    Value *StackTag =
+        ClGenerateTagsWithCalls ? nullptr : getStackBaseTag(EntryIRB);
+    Changed |= instrumentStack(AllocasToInstrument, RetVec, StackTag);
+  }
 
   for (auto Inst : ToInstrument)
     Changed |= instrumentMemAccess(Inst);
@@ -752,26 +864,26 @@
 }
 
 void HWAddressSanitizer::ShadowMapping::init(Triple &TargetTriple) {
-  const bool IsAndroid = TargetTriple.isAndroid();
-  const bool IsAndroidWithIfuncSupport =
-      IsAndroid && !TargetTriple.isAndroidVersionLT(21);
-
   Scale = kDefaultShadowScale;
-  const bool WithIfunc = ClWithIfunc.getNumOccurrences() > 0
-                             ? ClWithIfunc
-                             : IsAndroidWithIfuncSupport;
-
   if (ClMappingOffset.getNumOccurrences() > 0) {
     InGlobal = false;
+    InTls = false;
     Offset = ClMappingOffset;
   } else if (ClEnableKhwasan || ClInstrumentWithCalls) {
     InGlobal = false;
+    InTls = false;
     Offset = 0;
-  } else if (WithIfunc) {
+  } else if (ClWithIfunc) {
     InGlobal = true;
+    InTls = false;
+    Offset = kDynamicShadowSentinel;
+  } else if (ClWithTls) {
+    InGlobal = false;
+    InTls = true;
     Offset = kDynamicShadowSentinel;
   } else {
     InGlobal = false;
+    InTls = false;
     Offset = kDynamicShadowSentinel;
   }
 }
Index: llvm/trunk/test/Instrumentation/HWAddressSanitizer/alloca.ll
===================================================================
--- llvm/trunk/test/Instrumentation/HWAddressSanitizer/alloca.ll
+++ llvm/trunk/test/Instrumentation/HWAddressSanitizer/alloca.ll
@@ -1,8 +1,8 @@
 ; Test alloca instrumentation.
 ;
-; RUN: opt < %s -hwasan -S | FileCheck %s --check-prefixes=CHECK,DYNAMIC-SHADOW,NO-UAR-TAGS
+; RUN: opt < %s -hwasan -hwasan-with-ifunc=1 -S | FileCheck %s --check-prefixes=CHECK,DYNAMIC-SHADOW,NO-UAR-TAGS
 ; RUN: opt < %s -hwasan -hwasan-mapping-offset=0 -S | FileCheck %s --check-prefixes=CHECK,ZERO-BASED-SHADOW,NO-UAR-TAGS
-; RUN: opt < %s -hwasan -hwasan-uar-retag-to-zero=0 -S | FileCheck %s --check-prefixes=CHECK,DYNAMIC-SHADOW,UAR-TAGS
+; RUN: opt < %s -hwasan -hwasan-with-ifunc=1 -hwasan-uar-retag-to-zero=0 -S | FileCheck %s --check-prefixes=CHECK,DYNAMIC-SHADOW,UAR-TAGS
 
 target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
 target triple = "aarch64--linux-android"
Index: llvm/trunk/test/Instrumentation/HWAddressSanitizer/basic.ll
===================================================================
--- llvm/trunk/test/Instrumentation/HWAddressSanitizer/basic.ll
+++ llvm/trunk/test/Instrumentation/HWAddressSanitizer/basic.ll
@@ -1,7 +1,7 @@
 ; Test basic address sanitizer instrumentation.
 ;
-; RUN: opt < %s -hwasan -hwasan-recover=0 -S | FileCheck %s --check-prefixes=CHECK,ABORT,DYNAMIC-SHADOW
-; RUN: opt < %s -hwasan -hwasan-recover=1 -S | FileCheck %s --check-prefixes=CHECK,RECOVER,DYNAMIC-SHADOW
+; RUN: opt < %s -hwasan -hwasan-recover=0 -hwasan-with-ifunc=1 -hwasan-with-tls=0 -S | FileCheck %s --check-prefixes=CHECK,ABORT,DYNAMIC-SHADOW
+; RUN: opt < %s -hwasan -hwasan-recover=1 -hwasan-with-ifunc=1 -hwasan-with-tls=0 -S | FileCheck %s --check-prefixes=CHECK,RECOVER,DYNAMIC-SHADOW
 ; RUN: opt < %s -hwasan -hwasan-recover=0 -hwasan-mapping-offset=0 -S | FileCheck %s --check-prefixes=CHECK,ABORT,ZERO-BASED-SHADOW
 ; RUN: opt < %s -hwasan -hwasan-recover=1 -hwasan-mapping-offset=0 -S | FileCheck %s --check-prefixes=CHECK,RECOVER,ZERO-BASED-SHADOW
 
@@ -342,7 +342,6 @@
 define i8 @test_load_addrspace(i8 addrspace(256)* %a) sanitize_hwaddress {
 ; CHECK-LABEL: @test_load_addrspace(
 ; CHECK-NEXT: entry:
-; DYNAMIC-SHADOW: %.hwasan.shadow = call i64 asm "", "=r,0"([0 x i8]* @__hwasan_shadow)
 ; CHECK-NEXT: %[[B:[^ ]*]] = load i8, i8 addrspace(256)* %a
 ; CHECK-NEXT: ret i8 %[[B]]
 
Index: llvm/trunk/test/Instrumentation/HWAddressSanitizer/prologue.ll
===================================================================
--- llvm/trunk/test/Instrumentation/HWAddressSanitizer/prologue.ll
+++ llvm/trunk/test/Instrumentation/HWAddressSanitizer/prologue.ll
@@ -0,0 +1,88 @@
+; Test -hwasan-with-ifunc flag.
+;
+; RUN: opt -hwasan -S < %s | \
+; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-TLS,CHECK-HISTORY
+; RUN: opt -hwasan -S -hwasan-with-ifunc=0 -hwasan-with-tls=1 -hwasan-record-stack-history=1 < %s | \
+; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-TLS,CHECK-HISTORY
+; RUN: opt -hwasan -S -hwasan-with-ifunc=0 -hwasan-with-tls=1 -hwasan-record-stack-history=0 < %s | \
+; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-TLS,CHECK-NOHISTORY
+; RUN: opt -hwasan -S -hwasan-with-ifunc=0 -hwasan-with-tls=0 < %s | \
+; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-GLOBAL,CHECK-NOHISTORY
+; RUN: opt -hwasan -S -hwasan-with-ifunc=1  -hwasan-with-tls=0 < %s | \
+; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-IFUNC,CHECK-NOHISTORY
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64--linux-android22"
+
+; CHECK-IFUNC: @__hwasan_shadow = external global [0 x i8]
+; CHECK-NOIFUNC: @__hwasan_shadow_memory_dynamic_address = external global i64
+
+define i32 @test_load(i32* %a) sanitize_hwaddress {
+; First instrumentation in the function must be to load the dynamic shadow
+; address into a local variable.
+; CHECK-LABEL: @test_load
+; CHECK: entry:
+
+; CHECK-IFUNC:   %[[A:[^ ]*]] = call i64 asm "", "=r,0"([0 x i8]* @__hwasan_shadow)
+; CHECK-IFUNC:   add i64 %{{.*}}, %[[A]]
+
+; CHECK-GLOBAL: load i64, i64* @__hwasan_shadow_memory_dynamic_address
+
+; CHECK-TLS:   %[[A:[^ ]*]] = call i8* @llvm.thread.pointer()
+; CHECK-TLS:   %[[B:[^ ]*]] = getelementptr i8, i8* %[[A]], i32 64
+; CHECK-TLS:   %[[C:[^ ]*]] = bitcast i8* %[[B]] to i64*
+; CHECK-TLS:   %[[D:[^ ]*]] = load i64, i64* %[[C]]
+; CHECK-TLS:   %[[E:[^ ]*]] = or i64 %[[D]], 4294967295
+; CHECK-TLS:   = add i64 %[[E]], 1
+
+; "store i64" is only used to update stack history (this input IR intentionally does not use any i64)
+; W/o any allocas, the history is not updated, even if it is enabled explicitly with -hwasan-record-stack-history=1
+; CHECK-NOT: store i64
+
+; CHECK: ret i32
+
+entry:
+  %x = load i32, i32* %a, align 4
+  ret i32 %x
+}
+
+declare void @use(i32* %p)
+
+define void @test_alloca() sanitize_hwaddress {
+; First instrumentation in the function must be to load the dynamic shadow
+; address into a local variable.
+; CHECK-LABEL: @test_alloca
+; CHECK: entry:
+
+; CHECK-IFUNC:   %[[A:[^ ]*]] = call i64 asm "", "=r,0"([0 x i8]* @__hwasan_shadow)
+; CHECK-IFUNC:   add i64 %{{.*}}, %[[A]]
+
+; CHECK-GLOBAL: load i64, i64* @__hwasan_shadow_memory_dynamic_address
+
+; CHECK-TLS:   %[[A:[^ ]*]] = call i8* @llvm.thread.pointer()
+; CHECK-TLS:   %[[B:[^ ]*]] = getelementptr i8, i8* %[[A]], i32 64
+; CHECK-TLS:   %[[C:[^ ]*]] = bitcast i8* %[[B]] to i64*
+; CHECK-TLS:   %[[D:[^ ]*]] = load i64, i64* %[[C]]
+
+; CHECK-NOHISTORY-NOT: store i64
+
+; CHECK-HISTORY: %[[PTR:[^ ]*]] = inttoptr i64 %[[D]] to i64*
+; CHECK-HISTORY: store i64 %{{.*}}, i64* %[[PTR]]
+; CHECK-HISTORY: %[[D1:[^ ]*]] = ashr i64 %[[D]], 56
+; CHECK-HISTORY: %[[D2:[^ ]*]] = shl nuw nsw i64 %[[D1]], 12
+; CHECK-HISTORY: %[[D3:[^ ]*]] = xor i64 %[[D2]], -1
+; CHECK-HISTORY: %[[D4:[^ ]*]] = add i64 %[[D]], 8
+; CHECK-HISTORY: %[[D5:[^ ]*]] = and i64 %[[D4]], %[[D3]]
+; CHECK-HISTORY: store i64 %[[D5]], i64* %[[C]]
+
+; CHECK-TLS:   %[[E:[^ ]*]] = or i64 %[[D]], 4294967295
+; CHECK-TLS:   = add i64 %[[E]], 1
+
+; CHECK-NOHISTORY-NOT: store i64
+
+
+entry:
+  %x = alloca i32, align 4
+  call void @use(i32* %x)
+  ret void
+}
Index: llvm/trunk/test/Instrumentation/HWAddressSanitizer/with-ifunc.ll
===================================================================
--- llvm/trunk/test/Instrumentation/HWAddressSanitizer/with-ifunc.ll
+++ llvm/trunk/test/Instrumentation/HWAddressSanitizer/with-ifunc.ll
@@ -1,30 +0,0 @@
-; Test -hwasan-with-ifunc flag.
-;
-; RUN: opt -hwasan -S < %s | \
-; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-IFUNC
-; RUN: opt -hwasan -S -hwasan-with-ifunc=0 < %s | \
-; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-NOIFUNC
-; RUN: opt -hwasan -S -hwasan-with-ifunc=1 < %s | \
-; RUN:     FileCheck %s --check-prefixes=CHECK,CHECK-IFUNC
-
-target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
-target triple = "aarch64--linux-android22"
-
-; CHECK-IFUNC: @__hwasan_shadow = external global [0 x i8]
-; CHECK-NOIFUNC: @__hwasan_shadow_memory_dynamic_address = external global i64
-
-define i32 @test_load(i32* %a) sanitize_hwaddress {
-; First instrumentation in the function must be to load the dynamic shadow
-; address into a local variable.
-; CHECK-LABEL: @test_load
-; CHECK: entry:
-
-; CHECK-IFUNC:   %[[A:[^ ]*]] = call i64 asm "", "=r,0"([0 x i8]* @__hwasan_shadow)
-; CHECK-IFUNC:   add i64 %{{.*}}, %[[A]]
-
-; CHECK-NOIFUNC: load i64, i64* @__hwasan_shadow_memory_dynamic_address
-
-entry:
-  %x = load i32, i32* %a, align 4
-  ret i32 %x
-}