[asan] Added assembly functions for x86/amd64 asan.

These are runtime support functions for inline assembly instrumentation.

Patch by Yuri Gorshenin.

llvm-svn: 201402

4 files changed, 869 insertions, 0 deletions

diff --git a/compiler-rt/lib/asan/CMakeLists.txt b/compiler-rt/lib/asan/CMakeLists.txt
index 52eb8efe6ec..8f329531d59 100644
--- a/compiler-rt/lib/asan/CMakeLists.txt
+++ b/compiler-rt/lib/asan/CMakeLists.txt
@@ -50,6 +50,10 @@ endif()
 filter_available_targets(ASAN_SUPPORTED_ARCH
   x86_64 i386 powerpc64)
 
+set(ASAN_ASM_SOURCES asan_asm_instrumentation.S)
+set_source_files_properties(${ASAN_ASM_SOURCES} PROPERTIES LANGUAGE C)
+list(APPEND ASAN_SOURCES ${ASAN_ASM_SOURCES})
+
 # Compile ASan sources into an object library.
 if(APPLE)
   foreach(os ${SANITIZER_COMMON_SUPPORTED_DARWIN_OS})
diff --git a/compiler-rt/lib/asan/asan_asm_instrumentation.S b/compiler-rt/lib/asan/asan_asm_instrumentation.S
new file mode 100644
index 00000000000..7de45825c84
--- /dev/null
+++ b/compiler-rt/lib/asan/asan_asm_instrumentation.S
@@ -0,0 +1,451 @@
+// This file was generated by gen_asm_instrumentation.sh. Please, do not edit
+// manually.
+.section .text
+#if defined(__i386__)
+.globl __sanitizer_sanitize_store1
+__sanitizer_sanitize_store1:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushl %edx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  movb 0x20000000(%ecx), %cl
+  testb %cl, %cl
+  je .sanitize_store1_done
+  movl %eax, %edx
+  andl $0x7, %edx
+  movsbl %cl, %ecx
+  cmpl %ecx, %edx
+  jl .sanitize_store1_done
+  pushl %eax
+  call __asan_report_store1
+.sanitize_store1_done:
+  popfl
+  popl %edx
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_load1
+__sanitizer_sanitize_load1:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushl %edx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  movb 0x20000000(%ecx), %cl
+  testb %cl, %cl
+  je .sanitize_load1_done
+  movl %eax, %edx
+  andl $0x7, %edx
+  movsbl %cl, %ecx
+  cmpl %ecx, %edx
+  jl .sanitize_load1_done
+  pushl %eax
+  call __asan_report_load1
+.sanitize_load1_done:
+  popfl
+  popl %edx
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_store2
+__sanitizer_sanitize_store2:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushl %edx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  movb 0x20000000(%ecx), %cl
+  testb %cl, %cl
+  je .sanitize_store2_done
+  movl %eax, %edx
+  andl $0x7, %edx
+  incl %edx
+  movsbl %cl, %ecx
+  cmpl %ecx, %edx
+  jl .sanitize_store2_done
+  pushl %eax
+  call __asan_report_store2
+.sanitize_store2_done:
+  popfl
+  popl %edx
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_load2
+__sanitizer_sanitize_load2:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushl %edx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  movb 0x20000000(%ecx), %cl
+  testb %cl, %cl
+  je .sanitize_load2_done
+  movl %eax, %edx
+  andl $0x7, %edx
+  incl %edx
+  movsbl %cl, %ecx
+  cmpl %ecx, %edx
+  jl .sanitize_load2_done
+  pushl %eax
+  call __asan_report_load2
+.sanitize_load2_done:
+  popfl
+  popl %edx
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_store4
+__sanitizer_sanitize_store4:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushl %edx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  movb 0x20000000(%ecx), %cl
+  testb %cl, %cl
+  je .sanitize_store4_done
+  movl %eax, %edx
+  andl $0x7, %edx
+  addl $0x3, %edx
+  movsbl %cl, %ecx
+  cmpl %ecx, %edx
+  jl .sanitize_store4_done
+  pushl %eax
+  call __asan_report_store4
+.sanitize_store4_done:
+  popfl
+  popl %edx
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_load4
+__sanitizer_sanitize_load4:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushl %edx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  movb 0x20000000(%ecx), %cl
+  testb %cl, %cl
+  je .sanitize_load4_done
+  movl %eax, %edx
+  andl $0x7, %edx
+  addl $0x3, %edx
+  movsbl %cl, %ecx
+  cmpl %ecx, %edx
+  jl .sanitize_load4_done
+  pushl %eax
+  call __asan_report_load4
+.sanitize_load4_done:
+  popfl
+  popl %edx
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_store8
+__sanitizer_sanitize_store8:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  cmpb $0x0, 0x20000000(%ecx)
+  je .sanitize_store8_done
+  pushl %eax
+  call __asan_report_store8
+.sanitize_store8_done:
+  popfl
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_load8
+__sanitizer_sanitize_load8:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  cmpb $0x0, 0x20000000(%ecx)
+  je .sanitize_load8_done
+  pushl %eax
+  call __asan_report_load8
+.sanitize_load8_done:
+  popfl
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_store16
+__sanitizer_sanitize_store16:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  cmpw $0x0, 0x20000000(%ecx)
+  je .sanitize_store16_done
+  pushl %eax
+  call __asan_report_store16
+.sanitize_store16_done:
+  popfl
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+.globl __sanitizer_sanitize_load16
+__sanitizer_sanitize_load16:
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl $0x3, %ecx
+  cmpw $0x0, 0x20000000(%ecx)
+  je .sanitize_load16_done
+  pushl %eax
+  call __asan_report_load16
+.sanitize_load16_done:
+  popfl
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+#endif // defined(__i386__)
+#if defined(__x86_64__)
+.globl __sanitizer_sanitize_store1
+__sanitizer_sanitize_store1:
+  pushq %rax
+  pushq %rcx
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  movb 0x7fff8000(%rax), %al
+  test %al, %al
+  je .sanitize_store1_done
+  movl %edi, %ecx
+  andl $0x7, %ecx
+  movsbl %al, %eax
+  cmpl %eax, %ecx
+  jl .sanitize_store1_done
+  call __asan_report_store1
+.sanitize_store1_done:
+  popfq
+  popq %rcx
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_load1
+__sanitizer_sanitize_load1:
+  pushq %rax
+  pushq %rcx
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  movb 0x7fff8000(%rax), %al
+  test %al, %al
+  je .sanitize_load1_done
+  movl %edi, %ecx
+  andl $0x7, %ecx
+  movsbl %al, %eax
+  cmpl %eax, %ecx
+  jl .sanitize_load1_done
+  call __asan_report_load1
+.sanitize_load1_done:
+  popfq
+  popq %rcx
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_store2
+__sanitizer_sanitize_store2:
+  pushq %rax
+  pushq %rcx
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  movb 0x7fff8000(%rax), %al
+  test %al, %al
+  je .sanitize_store2_done
+  movl %edi, %ecx
+  andl $0x7, %ecx
+  incl %ecx
+  movsbl %al, %eax
+  cmpl %eax, %ecx
+  jl .sanitize_store2_done
+  call __asan_report_store2
+.sanitize_store2_done:
+  popfq
+  popq %rcx
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_load2
+__sanitizer_sanitize_load2:
+  pushq %rax
+  pushq %rcx
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  movb 0x7fff8000(%rax), %al
+  test %al, %al
+  je .sanitize_load2_done
+  movl %edi, %ecx
+  andl $0x7, %ecx
+  incl %ecx
+  movsbl %al, %eax
+  cmpl %eax, %ecx
+  jl .sanitize_load2_done
+  call __asan_report_load2
+.sanitize_load2_done:
+  popfq
+  popq %rcx
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_store4
+__sanitizer_sanitize_store4:
+  pushq %rax
+  pushq %rcx
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  movb 0x7fff8000(%rax), %al
+  test %al, %al
+  je .sanitize_store4_done
+  movl %edi, %ecx
+  andl $0x7, %ecx
+  addl $0x3, %ecx
+  movsbl %al, %eax
+  cmpl %eax, %ecx
+  jl .sanitize_store4_done
+  call __asan_report_store4
+.sanitize_store4_done:
+  popfq
+  popq %rcx
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_load4
+__sanitizer_sanitize_load4:
+  pushq %rax
+  pushq %rcx
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  movb 0x7fff8000(%rax), %al
+  test %al, %al
+  je .sanitize_load4_done
+  movl %edi, %ecx
+  andl $0x7, %ecx
+  addl $0x3, %ecx
+  movsbl %al, %eax
+  cmpl %eax, %ecx
+  jl .sanitize_load4_done
+  call __asan_report_load4
+.sanitize_load4_done:
+  popfq
+  popq %rcx
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_store8
+__sanitizer_sanitize_store8:
+  pushq %rax
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  cmpb $0x0, 0x7fff8000(%rax)
+  je .sanitize_store8_done
+  call __asan_report_store8
+.sanitize_store8_done:
+  popfq
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_load8
+__sanitizer_sanitize_load8:
+  pushq %rax
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  cmpb $0x0, 0x7fff8000(%rax)
+  je .sanitize_load8_done
+  call __asan_report_load8
+.sanitize_load8_done:
+  popfq
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_store16
+__sanitizer_sanitize_store16:
+  pushq %rax
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  cmpw $0x0, 0x7fff8000(%rax)
+  je .sanitize_store16_done
+  call __asan_report_store16
+.sanitize_store16_done:
+  popfq
+  popq %rax
+  ret
+.globl __sanitizer_sanitize_load16
+__sanitizer_sanitize_load16:
+  pushq %rax
+  pushfq
+  movq %rdi, %rax
+  shrq $0x3, %rax
+  cmpw $0x0, 0x7fff8000(%rax)
+  je .sanitize_load16_done
+  call __asan_report_load16
+.sanitize_load16_done:
+  popfq
+  popq %rax
+  ret
+#endif // defined(__x86_64__)
diff --git a/compiler-rt/lib/asan/scripts/gen_asm_instrumentation.sh b/compiler-rt/lib/asan/scripts/gen_asm_instrumentation.sh
new file mode 100755
index 00000000000..e423a53d81e
--- /dev/null
+++ b/compiler-rt/lib/asan/scripts/gen_asm_instrumentation.sh
@@ -0,0 +1,215 @@
+#!/bin/bash
+
+#===- lib/asan/scripts/gen_asm_instrumentation.sh -------------------------===#
+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+# Emit x86 instrumentation functions for asan.
+#
+#===-----------------------------------------------------------------------===#
+
+
+check() {
+  test $# -eq 2 || (echo "Incorrent number of arguments: $#" 1>&2 && exit 1)
+  case "$1" in
+    store) ;;
+     load) ;;
+        *) echo "Incorrect first argument: $1" 1>&2 && exit 1 ;;
+  esac
+  case "$2" in
+    [0-9]*) ;;
+         *) echo "Incorrect second argument: $2" 1>&2 && exit 1 ;;
+  esac
+}
+
+fname() {
+  check $1 $2
+  echo "__sanitizer_sanitize_$1$2"
+}
+
+flabel() {
+  check $1 $2
+  echo ".sanitize_$1$2_done"
+}
+
+freport() {
+  check $1 $2
+  echo "__asan_report_$1$2"
+}
+
+cat <<EOF
+// This file was generated by $(basename $0). Please, do not edit
+// manually.
+EOF
+
+echo ".section .text"
+
+echo "#if defined(__i386__)"
+
+# Functions for i386 1-, 2- and 4-byte accesses.
+for as in 1 2 4
+do
+  for at in store load
+  do
+cat <<EOF
+.globl $(fname $at $as)
+$(fname $at $as):
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushl %edx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl \$0x3, %ecx
+  movb 0x20000000(%ecx), %cl
+  testb %cl, %cl
+  je $(flabel $at $as)
+  movl %eax, %edx
+  andl \$0x7, %edx
+EOF
+
+    case $as in
+    1) ;;
+    2) echo '  incl %edx' ;;
+    4) echo '  addl $0x3, %edx' ;;
+    *) echo "Incorrect access size: $as" 1>&2; exit 1 ;;
+    esac
+
+cat <<EOF
+  movsbl %cl, %ecx
+  cmpl %ecx, %edx
+  jl $(flabel $at $as)
+  pushl %eax
+  call $(freport $at $as)
+$(flabel $at $as):
+  popfl
+  popl %edx
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+EOF
+  done
+done
+
+# Functions for i386 8- and 16-byte accesses.
+for as in 8 16
+do
+  for at in store load
+  do
+cat <<EOF
+.globl $(fname $at $as)
+$(fname $at $as):
+  pushl %ebp
+  movl %esp, %ebp
+  pushl %eax
+  pushl %ecx
+  pushfl
+  movl 8(%ebp), %eax
+  movl %eax, %ecx
+  shrl \$0x3, %ecx
+EOF
+
+    case ${as} in
+      8) echo '  cmpb $0x0, 0x20000000(%ecx)' ;;
+     16) echo '  cmpw $0x0, 0x20000000(%ecx)' ;;
+      *) echo "Incorrect access size: ${as}" 1>&2; exit 1 ;;
+    esac
+
+cat <<EOF
+  je $(flabel $at $as)
+  pushl %eax
+  call $(freport $at $as)
+$(flabel $at $as):
+  popfl
+  popl %ecx
+  popl %eax
+  movl %ebp, %esp
+  popl %ebp
+  ret
+EOF
+  done
+done
+
+echo "#endif // defined(__i386__)"
+
+echo "#if defined(__x86_64__)"
+
+# Functions for x86-64 1-, 2- and 4-byte accesses.
+for as in 1 2 4
+do
+  for at in store load
+  do
+cat <<EOF
+.globl $(fname $at $as)
+$(fname $at $as):
+  pushq %rax
+  pushq %rcx
+  pushfq
+  movq %rdi, %rax
+  shrq \$0x3, %rax
+  movb 0x7fff8000(%rax), %al
+  test %al, %al
+  je $(flabel $at $as)
+  movl %edi, %ecx
+  andl \$0x7, %ecx
+EOF
+
+    case ${as} in
+    1) ;;
+    2) echo '  incl %ecx' ;;
+    4) echo '  addl $0x3, %ecx' ;;
+    *) echo "Incorrect access size: ${as}" 1>&2; exit 1 ;;
+    esac
+
+cat <<EOF
+  movsbl %al, %eax
+  cmpl %eax, %ecx
+  jl $(flabel $at $as)
+  call $(freport $at $as)
+$(flabel $at $as):
+  popfq
+  popq %rcx
+  popq %rax
+  ret
+EOF
+  done
+done
+
+# Functions for x86-64 8- and 16-byte accesses.
+for as in 8 16
+do
+  for at in store load
+  do
+cat <<EOF
+.globl $(fname $at $as)
+$(fname $at $as):
+  pushq %rax
+  pushfq
+  movq %rdi, %rax
+  shrq \$0x3, %rax
+EOF
+
+    case ${as} in
+      8) echo '  cmpb $0x0, 0x7fff8000(%rax)' ;;
+     16) echo '  cmpw $0x0, 0x7fff8000(%rax)' ;;
+      *) echo "Incorrect access size: ${as}" 1>&2; exit 1 ;;
+    esac
+
+cat <<EOF
+  je $(flabel $at $as)
+  call $(freport $at $as)
+$(flabel $at $as):
+  popfq
+  popq %rax
+  ret
+EOF
+  done
+done
+echo "#endif // defined(__x86_64__)"
diff --git a/compiler-rt/lib/asan/tests/asan_test.cc b/compiler-rt/lib/asan/tests/asan_test.cc
index af7c471883b..587d85bea74 100644
--- a/compiler-rt/lib/asan/tests/asan_test.cc
+++ b/compiler-rt/lib/asan/tests/asan_test.cc
@@ -1244,3 +1244,202 @@ TEST(AddressSanitizer, pthread_getschedparam) {
   int res = pthread_getschedparam(pthread_self(), &policy, &param);
   ASSERT_EQ(0, res);
 }
+
+// Tests for __sanitizer_sanitize_(store|load)N functions in compiler-rt.
+
+#if defined(__x86_64__) || (defined(__i386__) && defined(__SSE2__))
+
+#include <emmintrin.h>
+
+namespace {
+
+template<typename T> void asm_write(T *ptr, T val);
+template<typename T> T asm_read(T *ptr);
+
+} // End of anonymous namespace
+
+#endif // defined(__x86_64__) || (defined(__i386__) && defined(__SSE2__))
+
+#if defined(__x86_64__)
+
+namespace {
+
+#define DECLARE_ASM_WRITE(Type, Size, Mov, Reg)        \
+template<> void asm_write<Type>(Type *ptr, Type val) { \
+  __asm__(                                             \
+    "leaq (%[ptr]), %%rdi  \n\t"                       \
+    "call __sanitizer_sanitize_store" Size "  \n\t"    \
+    Mov " %[val], (%[ptr])  \n\t"                      \
+    :                                                  \
+    : [ptr] "r" (ptr), [val] Reg (val)                 \
+    : "memory", "rdi"                                  \
+  );                                                   \
+}
+
+#define DECLARE_ASM_READ(Type, Size, Mov, Reg)     \
+template<> Type asm_read<Type>(Type *ptr) {        \
+  Type res;                                        \
+  __asm__(                                         \
+    "leaq (%[ptr]), %%rdi  \n\t"                   \
+    "call __sanitizer_sanitize_load" Size "  \n\t" \
+    Mov " (%[ptr]), %[res]  \n\t"                  \
+    : [res] Reg (res)                              \
+    : [ptr] "r" (ptr)                              \
+    : "memory", "rdi"                              \
+  );                                               \
+  return res;                                      \
+}
+
+DECLARE_ASM_WRITE(U8, "8", "movq", "r");
+DECLARE_ASM_READ(U8, "8", "movq", "=r");
+
+} // End of anonymous namespace
+
+#endif // defined(__x86_64__)
+
+#if defined(__i386__) && defined(__SSE2__)
+
+namespace {
+
+#define DECLARE_ASM_WRITE(Type, Size, Mov, Reg)        \
+template<> void asm_write<Type>(Type *ptr, Type val) { \
+  __asm__(                                             \
+    "leal (%[ptr]), %%eax  \n\t"                       \
+    "pushl %%eax  \n\t"                                \
+    "call __sanitizer_sanitize_store" Size "  \n\t"    \
+    "popl %%eax  \n\t"                                 \
+    Mov " %[val], (%[ptr])  \n\t"                      \
+    :                                                  \
+    : [ptr] "r" (ptr), [val] Reg (val)                 \
+    : "memory", "eax", "esp"                           \
+  );                                                   \
+}
+
+#define DECLARE_ASM_READ(Type, Size, Mov, Reg)     \
+template<> Type asm_read<Type>(Type *ptr) {        \
+  Type res;                                        \
+  __asm__(                                         \
+    "leal (%[ptr]), %%eax  \n\t"                   \
+    "pushl %%eax  \n\t"                            \
+    "call __sanitizer_sanitize_load" Size "  \n\t" \
+    "popl %%eax  \n\t"                             \
+    Mov " (%[ptr]), %[res]  \n\t"                  \
+    : [res] Reg (res)                              \
+    : [ptr] "r" (ptr)                              \
+    : "memory", "eax", "esp"                       \
+  );                                               \
+  return res;                                      \
+}
+
+template<> void asm_write<U8>(U8 *ptr, U8 val) {
+  __asm__(
+    "leal (%[ptr]), %%eax  \n\t"
+    "pushl %%eax  \n\t"
+    "call __sanitizer_sanitize_store8  \n\t"
+    "popl %%eax  \n\t"
+    "movl (%[val]), %%eax  \n\t"
+    "movl %%eax, (%[ptr])  \n\t"
+    "movl 0x4(%[val]), %%eax  \n\t"
+    "movl %%eax, 0x4(%[ptr])  \n\t"
+    :
+    : [ptr] "r" (ptr), [val] "r" (&val)
+    : "memory", "eax", "esp"
+  );
+}
+
+template<> U8 asm_read(U8 *ptr) {
+  U8 res;
+  __asm__(
+    "leal (%[ptr]), %%eax  \n\t"
+    "pushl %%eax  \n\t"
+    "call __sanitizer_sanitize_load8  \n\t"
+    "popl  %%eax  \n\t"
+    "movl (%[ptr]), %%eax  \n\t"
+    "movl %%eax, (%[res])  \n\t"
+    "movl 0x4(%[ptr]), %%eax  \n\t"
+    "movl %%eax, 0x4(%[res])  \n\t"
+    :
+    : [ptr] "r" (ptr), [res] "r" (&res)
+    : "memory", "eax", "esp"
+  );
+  return res;
+}
+
+} // End of anonymous namespace
+
+#endif  // defined(__i386__) && defined(__SSE2__)
+
+#if defined(__x86_64__) || (defined(__i386__) && defined(__SSE2__))
+
+namespace {
+
+DECLARE_ASM_WRITE(U1, "1", "movb", "r");
+DECLARE_ASM_WRITE(U2, "2", "movw", "r");
+DECLARE_ASM_WRITE(U4, "4", "movl", "r");
+DECLARE_ASM_WRITE(__m128i, "16", "movaps", "x");
+
+DECLARE_ASM_READ(U1, "1", "movb", "=r");
+DECLARE_ASM_READ(U2, "2", "movw", "=r");
+DECLARE_ASM_READ(U4, "4", "movl", "=r");
+DECLARE_ASM_READ(__m128i, "16", "movaps", "=x");
+
+template<typename T> void TestAsmWrite(const char *DeathPattern) {
+  T *buf = new T;
+  EXPECT_DEATH(asm_write(&buf[1], static_cast<T>(0)), DeathPattern);
+  T var = 0x12;
+  asm_write(&var, static_cast<T>(0x21));
+  ASSERT_EQ(static_cast<T>(0x21), var);
+  delete buf;
+}
+
+template<> void TestAsmWrite<__m128i>(const char *DeathPattern) {
+  char *buf = new char[16];
+  char *p = buf + 16;
+  if (((uintptr_t) p % 16) != 0)
+    p = buf + 8;
+  assert(((uintptr_t) p % 16) == 0);
+  __m128i val = _mm_set1_epi16(0x1234);
+  EXPECT_DEATH(asm_write<__m128i>((__m128i*) p, val), DeathPattern);
+  __m128i var = _mm_set1_epi16(0x4321);
+  asm_write(&var, val);
+  ASSERT_EQ(0x1234, _mm_extract_epi16(var, 0));
+  delete [] buf;
+}
+
+template<typename T> void TestAsmRead(const char *DeathPattern) {
+  T *buf = new T;
+  EXPECT_DEATH(asm_read(&buf[1]), DeathPattern);
+  T var = 0x12;
+  ASSERT_EQ(static_cast<T>(0x12), asm_read(&var));
+  delete buf;
+}
+
+template<> void TestAsmRead<__m128i>(const char *DeathPattern) {
+  char *buf = new char[16];
+  char *p = buf + 16;
+  if (((uintptr_t) p % 16) != 0)
+    p = buf + 8;
+  assert(((uintptr_t) p % 16) == 0);
+  EXPECT_DEATH(asm_read<__m128i>((__m128i*) p), DeathPattern);
+  __m128i val = _mm_set1_epi16(0x1234);
+  ASSERT_EQ(0x1234, _mm_extract_epi16(asm_read(&val), 0));
+  delete [] buf;
+}
+
+} // End of anonymous namespace
+
+TEST(AddressSanitizer, asm_rw) {
+  TestAsmWrite<U1>("WRITE of size 1");
+  TestAsmWrite<U2>("WRITE of size 2");
+  TestAsmWrite<U4>("WRITE of size 4");
+  TestAsmWrite<U8>("WRITE of size 8");
+  TestAsmWrite<__m128i>("WRITE of size 16");
+
+  TestAsmRead<U1>("READ of size 1");
+  TestAsmRead<U2>("READ of size 2");
+  TestAsmRead<U4>("READ of size 4");
+  TestAsmRead<U8>("READ of size 8");
+  TestAsmRead<__m128i>("READ of size 16");
+}
+
+#endif // defined(__x86_64__) || (defined(__i386__) && defined(__SSE2__))