diff --git a/.gitignore b/.gitignore
index d40845a..42328f2 100644
--- a/.gitignore
+++ b/.gitignore
@@ -143,3 +143,5 @@
 /test-driver
 /thread_dealloc_unittest
 /thread_dealloc_unittest.exe
+/unwind_bench
+/unwind_bench.exe
diff --git a/Makefile.am b/Makefile.am
index 4bd3107..0aeb77d 100755
--- a/Makefile.am
+++ b/Makefile.am
@@ -874,6 +874,12 @@ malloc_bench_shared_full_CXXFLAGS = $(PTHREAD_CFLAGS) $(AM_CXXFLAGS) $(NO_BUILTI
 malloc_bench_shared_full_LDFLAGS = $(PTHREAD_CFLAGS) $(TCMALLOC_FLAGS)
 malloc_bench_shared_full_LDADD = librun_benchmark.la libtcmalloc.la $(PTHREAD_LIBS)
 
+noinst_PROGRAMS += unwind_bench
+unwind_bench_SOURCES = benchmark/unwind_bench.cc benchmark/getcontext_light.cc
+unwind_bench_CXXFLAGS = $(PTHREAD_CFLAGS) $(AM_CXXFLAGS) $(NO_BUILTIN_CXXFLAGS)
+unwind_bench_LDFLAGS = $(PTHREAD_CFLAGS) $(TCMALLOC_FLAGS)
+unwind_bench_LDADD = librun_benchmark.la libtcmalloc.la $(PTHREAD_LIBS)
+
 endif WITH_HEAP_PROFILER_OR_CHECKER
 
 binary_trees_SOURCES = benchmark/binary_trees.cc
diff --git a/benchmark/getcontext_light.cc b/benchmark/getcontext_light.cc
new file mode 100644
index 0000000..e8ff8b3
--- /dev/null
+++ b/benchmark/getcontext_light.cc
@@ -0,0 +1,60 @@
+#include <ucontext.h>
+#include <stddef.h>
+
+extern "C" void getcontext_light(ucontext_t *ctx);
+
+// clang's built-in asm cannot handle .set directives
+#if defined(__GNUC__) && !defined(__llvm__) && defined(__x86_64) && defined(_LP64)
+
+#define R(r) offsetof(ucontext_t, uc_mcontext.gregs[r])
+
+static __attribute__((used))
+void getcontext_tramp(ucontext_t *ctx) {
+	__asm__ __volatile__(".set oRBX, %c0\n"
+			     ".set oRBP, %c1\n"
+			     ".set oR12, %c2\n"
+			     ".set oR13, %c3\n"
+			     ".set oR14, %c4\n"
+			     ".set oR15, %c5\n"
+			     ".set oRIP, %c6\n"
+			     ".set oRSP, %c7\n"
+			     : :
+			       "p" (R(REG_RBX)),
+			       "p" (R(REG_RBP)),
+			       "p" (R(REG_R12)),
+			       "p" (R(REG_R13)),
+			       "p" (R(REG_R14)),
+			       "p" (R(REG_R15)),
+			       "p" (R(REG_RIP)),
+			       "p" (R(REG_RSP)));
+	getcontext_light(ctx);
+}
+
+__asm__(".pushsection .text; .globl getcontext_light\n"
+	".type getcontext_light, @function\n"
+"getcontext_light:\n"
+	"\t.cfi_startproc\n"
+	"\tmovq     %rbx, oRBX(%rdi)\n"
+	"\tmovq     %rbp, oRBP(%rdi)\n"
+	"\tmovq     %r12, oR12(%rdi)\n"
+	"\tmovq     %r13, oR13(%rdi)\n"
+	"\tmovq     %r14, oR14(%rdi)\n"
+	"\tmovq     %r15, oR14(%rdi)\n"
+
+	"\tmovq     (%rsp), %rcx\n"
+	"\tmovq     %rcx, oRIP(%rdi)\n"
+	"\tleaq     8(%rsp), %rcx\n"                /* Exclude the return address.  */
+	"\tmovq     %rcx, oRSP(%rdi)\n"
+	"\tret\n"
+	".cfi_endproc\n"
+	".size getcontext_light, .-getcontext_light\n"
+	".popsection\n"
+	);
+
+#else
+
+extern "C" void getcontext_light(ucontext_t *ctx) {
+	getcontext(ctx);
+}
+
+#endif
diff --git a/benchmark/unwind_bench.cc b/benchmark/unwind_bench.cc
new file mode 100644
index 0000000..f919fc0
--- /dev/null
+++ b/benchmark/unwind_bench.cc
@@ -0,0 +1,89 @@
+// -*- Mode: C++; c-basic-offset: 2; indent-tabs-mode: nil -*-
+#include "config.h"
+
+#include "base/basictypes.h"
+#include "gperftools/stacktrace.h"
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdint.h>
+#include "sys/ucontext.h"
+#if HAVE_LIBUNWIND_H
+#include <libunwind.h>
+#endif
+
+#include "run_benchmark.h"
+
+extern "C" void getcontext_light(ucontext_t *ctx);
+
+#define MAX_FRAMES 1024
+static void *frames[MAX_FRAMES];
+
+enum measure_mode {
+  MODE_NOOP,
+  MODE_WITH_CONTEXT,
+  MODE_WITHOUT_CONTEXT
+};
+
+static int ATTRIBUTE_NOINLINE measure_unwind(int maxlevel, int mode) {
+  int n;
+
+  if (mode == MODE_NOOP)
+    return 0;
+
+  if (mode == MODE_WITH_CONTEXT) {
+    ucontext_t uc;
+    getcontext_light(&uc);
+    n = GetStackTraceWithContext(frames, MAX_FRAMES, 0, &uc);
+  } else {
+    n = GetStackTrace(frames, MAX_FRAMES, 0);
+  }
+  if (n < maxlevel) {
+    fprintf(stderr, "Expected at least %d frames, got %d\n", maxlevel, n);
+    abort();
+  }
+  return 0;
+}
+
+static int ATTRIBUTE_NOINLINE frame_forcer(int rv) {
+  return rv;
+}
+
+static int ATTRIBUTE_NOINLINE f1(int level, int maxlevel, int mode) {
+  if (level == maxlevel)
+    return frame_forcer(measure_unwind(maxlevel, mode));
+  return frame_forcer(f1(level + 1, maxlevel, mode));
+}
+
+static void bench_unwind_no_op(long iterations, uintptr_t param) {
+  do {
+    f1(0, param, MODE_NOOP);
+    iterations -= param;
+  } while (iterations > 0);
+}
+
+static void bench_unwind_context(long iterations, uintptr_t param) {
+  do {
+    f1(0, param, MODE_WITH_CONTEXT);
+    iterations -= param;
+  } while (iterations > 0);
+}
+
+static void bench_unwind_no_context(long iterations, uintptr_t param) {
+  do {
+    f1(0, param, MODE_WITHOUT_CONTEXT);
+    iterations -= param;
+  } while (iterations > 0);
+}
+
+int main(void) {
+  report_benchmark("unwind_no_op", bench_unwind_no_op, 100);
+  report_benchmark("unwind_context", bench_unwind_context, 100);
+  report_benchmark("unwind_no_context", bench_unwind_no_context, 100);
+
+//// TODO: somehow this fails at linking step. Figure out why this is missing
+// #if HAVE_LIBUNWIND_H
+//   unw_set_caching_policy(unw_local_addr_space, UNW_CACHE_PER_THREAD);
+// #endif
+  return 0;
+}