memset_explicit

A custom memset() implementation to force bypass GCC/Clang optimization

Supported Compiers

The code was tested under the following compilers:

• ARMCC
• GCC
• GCC-ARM
• Clang (x86, arm64, armv7, armv7s)

Demo

int main() {
  char buffer[4];
  buffer[0] = 'A';
  buffer[1] = 'B';
  buffer[2] = 'C';
  buffer[3] = 'D';
  /*
   * memset((unsigned char*)buffer, 0x00, sizeof(buffer));
   */
  memset_explicit((unsigned char*)buffer, 0x00, sizeof(buffer));
  return 0;
}

Typical memset with -O3 optimization produces the following assembly for the code above, clearly shorting out the memset call. Here's how the compiled assembly looks like after compiling it with GCC/Clang:

gdb$ Dump of assembler code for function main:
   0x00001ff0 <+0>:     push   %ebp
   0x00001ff1 <+1>:     mov    %esp,%ebp
   0x00001ff3 <+3>:     xor    %eax,%eax
   0x00001ff5 <+5>:     pop    %ebp
   0x00001ff6 <+6>:     ret
End of assembler dump.

As can be seen the memset call is "optimized" by GCC/Clang's optimizer.

This can be problematic, even dangerous, if the memset was critically needed to erase memory locations containting privacy sensitive data. These are the situations where memset_explicit(..) becomes useful. The same code snippet above compiled with -O3 produces the following assembly with memset_explicit(..) forcing the deletion of the buffer as the programmer have had intended.

gdb$ Dump of assembler code for function main:
   0x00001fd0 <+0>:     push   %ebp
   0x00001fd1 <+1>:     mov    %esp,%ebp
   0x00001fd3 <+3>:     sub    $0x18,%esp
   0x00001fd6 <+6>:     movl   $0x44434241,-0x4(%ebp)
   0x00001fdd <+13>:    lea    -0x4(%ebp),%eax
   0x00001fe0 <+16>:    mov    %eax,(%esp)
   0x00001fe3 <+19>:    movl   $0x4,0x8(%esp)
   0x00001feb <+27>:    movl   $0x0,0x4(%esp)
   0x00001ff3 <+35>:    call   0x1fb0 <memset_explicit>
   0x00001ff8 <+40>:    xor    %eax,%eax
   0x00001ffa <+42>:    add    $0x18,%esp
   0x00001ffd <+45>:    pop    %ebp
   0x00001ffe <+46>:    ret
End of assembler dump.