Search code examples
cgcclinkercompiler-optimization

Behaviour of -O and -wrap


I am trying to mock a function using -wrap=symbol for my unit testing and I find different behavior based on the GCC optimization flags passed

Consider the below example program

#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>

void * __wrap_malloc(size_t sz)
{
   printf("Called %s", __func__);
   return (void *) 0xDEADBEEF;
}

int main()
{
    void *ptr;
    ptr = malloc(10);
    assert(0xDEADBEEF == (unsigned int)(uintptr_t)(ptr));
    return 0;
}

When using optimization

❯ gcc -Os -Wl,-wrap=malloc 1.c -o opt

❯ ./opt
opt: 1.c:16: main: Assertion `0xDEADBEEF == (unsigned int)(uintptr_t)(ptr)' failed.
[1]    12546 abort      ./opt

Without optimization

❯ gcc  -Wl,-wrap=malloc 1.c -o without_opt
❯ ./without_opt
Called __wrap_malloc

I tried disassembling the executable

With optimization

   0000000000400450 <main>:
     400450:   48 83 ec 08             sub    $0x8,%rsp
     400454:   b9 20 06 40 00          mov    $0x400620,%ecx
     400459:   ba 10 00 00 00          mov    $0x10,%edx
     40045e:   be ea 05 40 00          mov    $0x4005ea,%esi
     400463:   bf ee 05 40 00          mov    $0x4005ee,%edi
     400468:   e8 d3 ff ff ff          callq  400440 <__assert_fail@plt>
     40046d:   0f 1f 00                nopl   (%rax)

Without optimization

   000000000040053e <main>:
     40053e:   55                      push   %rbp
     40053f:   48 89 e5                mov    %rsp,%rbp
     400542:   48 83 ec 10             sub    $0x10,%rsp
     400546:   bf 0a 00 00 00          mov    $0xa,%edi
     40054b:   e8 c7 ff ff ff          callq  400517 <__wrap_malloc>
     400550:   48 89 45 f8             mov    %rax,-0x8(%rbp)
     400554:   48 8b 45 f8             mov    -0x8(%rbp),%rax
     400558:   3d ef be ad de          cmp    $0xdeadbeef,%eax
     40055d:   74 19                   je     400578 <main+0x3a>
     40055f:   b9 4e 06 40 00          mov    $0x40064e,%ecx
     400564:   ba 10 00 00 00          mov    $0x10,%edx
     400569:   be 0a 06 40 00          mov    $0x40060a,%esi
     40056e:   bf 10 06 40 00          mov    $0x400610,%edi
     400573:   e8 c8 fe ff ff          callq  400440 <__assert_fail@plt>
     400578:   b8 00 00 00 00          mov    $0x0,%eax
     40057d:   c9                      leaveq
     40057e:   c3                      retq
     40057f:   90                      nop

There is no call to wrap_malloc. It looks like it got optimized out. Then I tried adding a print statement

   int main()
   {
       void *ptr;
       ptr = malloc(10);
 >>    printf("0x%x\n", (uintptr_t)ptr);
       assert(0xDEADBEEF == (unsigned int)(uintptr_t)(ptr));
       return 0;
   }


❯ gcc -Os -Wl,-wrap=malloc 1.c -o print_opt
❯ ./print_opt
Called __wrap_malloc0xdeadbeef
print_opt: 1.c:17: main: Assertion `0xDEADBEEF == (unsigned int)(uintptr_t)(ptr)' failed.
[1]    31575 abort      ./print_opt

Now I see the wrapped function getting called. But the assertion failed

   0000000000400450 <main>:
     400450:   48 83 ec 08             sub    $0x8,%rsp
     400454:   bf 0a 00 00 00          mov    $0xa,%edi
     400459:   e8 f9 00 00 00          callq  400557 <__wrap_malloc>
     40045e:   bf 0a 06 40 00          mov    $0x40060a,%edi
     400463:   48 89 c6                mov    %rax,%rsi
     400466:   31 c0                   xor    %eax,%eax
     400468:   e8 c3 ff ff ff          callq  400430 <printf@plt>
     40046d:   b9 48 06 40 00          mov    $0x400648,%ecx
     400472:   ba 11 00 00 00          mov    $0x11,%edx
     400477:   be 10 06 40 00          mov    $0x400610,%esi
     40047c:   bf 14 06 40 00          mov    $0x400614,%edi
     400481:   e8 ba ff ff ff          callq  400440 <__assert_fail@plt>
     400486:   66 2e 0f 1f 84 00 00    nopw   %cs:0x0(%rax,%rax,1)
     40048d:   00 00 00

Could some one help me understand why/how I am seeing this behavior?. Thanks in advance


Solution

  • With gcc, malloc() is treated as what's called a "built in" function that the optimizer is aware of and handles specially (It can make certain assumptions about its return value, for example, which lead it to automatically failing the assertion without needing a runtime test of it or even calling malloc() in the first place). If you add -fno-builtin-malloc to your compilation options, that special treatment is disabled:

    $ gcc -Os -fno-builtin-malloc -Wl,-wrap=malloc 1.c -o opt
    $ ./opt
    Called __wrap_malloc
    

    Side note: I added a trailing newline to your printf() output, because a program that produces textual output without one can cause odd effects in the calling shell - without it, the line was overwritten by my shell's prompt, for example. Or you might get a trailing % printed, or other behavior.