test fails because unhandled alignment trap

[Jani Monoses]

Binary package hint: vowpal-wabbit

The package FTBFS on armel. It builds actually but the testsuite fails.

This is in dmesg

[ 4806.357452] Alignment trap: not handling instruction ecb08a01 at [<00014658>]
[ 4806.365600] Unhandled fault: alignment exception (0x011) at 0x01563321

This being a numerically intensive library it is compiled in Debian with -ffast-math. On armel this makes gcc emit some vector instructions and which are not handled by the kernel's alignment trap handler and the test crashes on a vldmia

  0x00014658 <read_cached_features(parser*, void*)+312>: b0 ec 01 8a vldmiane r0!, {s16}

The C code that is compiled here uses a char * that is used to access data, so maybe that is the source of the problem.

for (;c!= end;)
        {
          feature f = {1., 0};
          size_t temp = f.weight_index;
          c = run_len_decode(c,temp);
          f.weight_index = temp;
          if (f.weight_index & neg_1)
            f.x = -1.;
          else if (f.weight_index & general) {
              f.x = *(float *)c;
              c += sizeof(float); // PC at crash
            }
          *our_sum_feat_sq += f.x*f.x;
          size_t diff = f.weight_index >> 2;
          int32_t s_diff = ZigZagDecode(diff);
          if (s_diff < 0)
            ae->sorted = false;
          f.weight_index = last + s_diff;
          last = f.weight_index;
          f.weight_index = f.weight_index & mask;
          push(*ours, f);
        }

[Launchpad Janitor (janitor)]

This bug was fixed in the package vowpal-wabbit - 5.0-1ubuntu1

---------------
vowpal-wabbit (5.0-1ubuntu1) natty; urgency=low

  * debian/patches/arm_alignment_fault.patch: Work around an ARM
    alignment trap caused by accessing unaligned float values.
    Thanks to Nicolas Pitre for suggesting the trick of casting
    to a packed struct containing a single float value.
    Fixes armel FTBFS (LP: #719305)
 -- Jani Monoses <email address hidden> Tue, 15 Feb 2011 21:57:50 +0200

[Jani Monoses (jani)]

BTW it was not ffast-math, even without any options gcc would emit VFP instructions.