------- Comment #7 from rguenth at gcc dot gnu dot org 2007-07-06 14:58 -------
We use store_field() to a bitpos of 80 with a bitsize of 48 to a mem:BLK target
which in turn calls store_expr on an adjusted mem which calls
temp = expand_expr_real (exp, tmp_target, GET_MODE (target),
(call_param_p
? EXPAND_STACK_PARM : EXPAND_NORMAL),
&alt_rtl);
that in turn creates this crapload of rtl via
else if (TYPE_MODE (TREE_TYPE (exp)) == BLKmode)
{
target = copy_blkmode_from_reg (target, valreg, TREE_TYPE (exp));
and
/* Copy the structure BITSIZE bites at a time.
We could probably emit more efficient code for machines which do not use
strict alignment, but it doesn't seem worth the effort at the current
time. */
not worth the effort!? uh :/
And the problem seems to be that we do store_bit_field of 1 byte at a time
but are using word_mode for it. At
(mem/s/j:DI (plus:DI (reg/f:DI 60)
(const_int 10 [0xa])) [0 <variable>.dRID+0 S8 A16])
storing with word_mode will write beyond the structure bounds though.
(store_bit_field will in turn use extraction and write-back to store each
of the bytes, so we will be better off using it directly for quantities
less than word_mode).
To summarize: copy_blkmode_from_reg () doesn't deal with a target size
that is less than the size of word_mode (or even a size that is not
a multiple of the size of word_mode).
The fix is either to make copy_blkmode_from_reg () handle this case or not
use that, but store_bit_field in the caller.
Of course maybe the bug is also that the type of the call expression has
BLKmode but the return value is passed in (reg:DI ax).
And of course I'm lost here ;)
--
rguenth at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |hubicka at gcc dot gnu dot
| |org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31309
