Testcase for pr37868 is broken; how shall we fix it?

[Mark Mitchell <mark at codesourcery dot com>]

The test case in gcc.dg/torture/pr37868.c is clearly broken, on many
hardware platforms.  It uses an "unsigned int *" to access a variable
whose alignment is not guaranteed.  Depending on the hardware that might
work, generate SIGBUS, or just result in bogus results.

  struct X {
    unsigned char pad : 4;
    unsigned int a : 32;
    unsigned int b : 24;
    unsigned int c : 6;
  } __attribute__((packed));

  int main (void)
  {
    struct X x;
    ...
    bad_bits = ((unsigned int)-1) ^ *(1+(unsigned int *) &x);
  }

The test also doesn't zero out x, before setting some of the fields, so
could well result in false passes, even on platforms without the
alignment problem.

We've already got a dg-skip-if for SPARC in the test.  But, in general,
this test is going to fail on other strict alignment platforms as well.
 ARM is an example; on ARM, this is a 9-byte structure with 1-byte
alignment and can end up with 1-byte alignment on the stack.

We don't have a strict-alignment check in target-supports.  I could add
one, hard-coding a list of platforms that are strict-alignment targets,
since there's no obvious way to probe at run-time.  (You can't guarantee
that you'll get a signal.)  I could just hard-code more platforms to
skip into the test.  I could fix the test to use char to read from the
unaligned memory location -- but it appears that the intent of the test
case was specifically to use this non-portable construct.

Recommendations?

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713