------- Comment #2 from rguenth at gcc dot gnu dot org 2010-04-13 11:20 -------
Well. There is
pretmp.2458_440 = c[2]{lb: 0 sz: 8};
...
c[2] = D.84046_29;
in a possibly dead code region dominated by
if (pretmp.2460_55 > 16)
goto <bb 54>;
else
goto <bb 57>;
where pretmp.2460_55 is (unsigned int)mode_size[mode_10(D)]. Thus,
the loop is peeled completely.
As we have
gcc_assert (j <= 2 * HOST_BITS_PER_WIDE_INT);
if (ch)
ch = (unsigned char) str[i];
c[j / HOST_BITS_PER_WIDE_INT] |= ch << (j % HOST_BITS_PER_WIDE_INT);
you can see that for j == 2 * HOST_BITS_PER_WIDE_INT we end up
accessing c[2]. In fact the assert looks wrong just for that reason.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43737
