[Bug c++/66730] New: Optimizer seems to make incorrect assumptions about function alignment

["kustermann.martin at gmail dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66730

            Bug ID: 66730
           Summary: Optimizer seems to make incorrect assumptions about
                    function alignment
           Product: gcc
           Version: 4.8.4
            Status: UNCONFIRMED
          Severity: major
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: kustermann.martin at gmail dot com
  Target Milestone: ---

Here is a reproduction of the problem

$ cat repro.cc
#include <stdio.h>

void Foobar()  { }

int main(int argc, char** argv) {
  unsigned long lowest_bit = reinterpret_cast<unsigned long>(&Foobar) & 1;
  printf("Lowest bit of %p is %ld\n", &Foobar, lowest_bit);
  return 0;
}

I expect this to print the address of the function and the lowest bit of that
address. Despite this assumption, the lowest bit is wrong:

$ g++ -O1 -o repro repro.cc
$ ./repro
Lowest bit of 0x40052d is 0

$ python
> bin(0x40052d)
'0b10000000000010100101101'

So the function is not 2-byte aligned, but the lowest bit gets reported as
being 0.
=> Either the functions must be guaranteed to have a proper alignment or the
optimizer cannot make alignment assumptions.

$ objdump -d repro
...
000000000040055f <main>:
  40055f:       48 83 ec 08             sub    $0x8,%rsp
  400563:       b9 00 00 00 00          mov    $0x0,%ecx
  400568:       ba 5d 05 40 00          mov    $0x40055d,%edx
  40056d:       be 14 06 40 00          mov    $0x400614,%esi
  400572:       bf 01 00 00 00          mov    $0x1,%edi
  400577:       b8 00 00 00 00          mov    $0x0,%eax
  40057c:       e8 df fe ff ff          callq  400460 <__printf_chk@plt>
  400581:       b8 00 00 00 00          mov    $0x0,%eax
  400586:       48 83 c4 08             add    $0x8,%rsp
  40058a:       c3                      retq   
  40058b:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
...


Beginning with optimization level '-O2' the problem seems to go away, since the
function is then properly aligned (not sure if this is guaranteed or not).

$ g++ --version
g++ (Ubuntu 4.8.4-2ubuntu1~14.04) 4.8.4
Copyright (C) 2013 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.


For reference with clang, this produces
$ clang repro.cc -o repro
$ ./repro
Lowest bit of 0x400530 is 0
$ clang --version
clang version 3.7.0 (trunk 239765)
Target: x86_64-unknown-linux-gnu
Thread model: posix



The workaround for us at the moment is to use an attribute:

void __attribute__((aligned(4))) Foobar() {}