On 11/11/13 07:32, Ulrich Weigand wrote:
Hello,
when implementing the new ABI for powerpc64le-linux, I ran into an assertion
failure in store_unaligned_arguments_into_pseudos:
gcc_assert (args[i].partial % UNITS_PER_WORD == 0);
This can happen in the new ABI since we pass "homogeneous structures"
consisting of soleley floating point elements of the same type in
floating-point registers until they run out, and the rest of the
structure in memory. If the structure member type is a 4-byte float,
and we only have an odd number of floating point registers available,
then args[i].partial can legitimately end up not being a multiple
of UNITS_PER_WORD (i.e. 8 on the platform).
Now, there are a number of similar checks that args[i].partial is
aligned elsewhere in calls.c and functions.c. But for all of those
the logic is: if args[i].reg is a REG, then args[i].partial must be
a multiple of the word size; but if args[i].regs is a PARALLEL, then
args[i].partial can be any arbitrary value. In the powerpc64le-linux
use case, the back-end always generates PARALLEL in this situation,
so it seemed the middle-end ought to support this -- and it does,
except for this one case.
However, looking more closely, it seems store_unaligned_arguments_into_pseudos
is not really useful for PARALLEL arguments in the first place. What this
routine does is load arguments into args[i].aligned_regs. But if we have
an argument where args[i].reg is a PARALLEL, args[i].aligned_regs will in
fact never be used later on at all! Instead, PARALLEL will always be
handled directly via emit_group_move (in load_register_parameters), so
the code generated by store_unaligned_arguments_into_pseudos for such cases
is simply dead anyway.
Thus, I'd suggest to simply have store_unaligned_arguments_into_pseudos
skip PARALLEL arguments.
Tested on powerpc64-linux and powerpc64le-linux.
OK for mainline?
Bye,
Ulrich
ChangeLog:
2013-11-11 Ulrich Weigand <ulrich.weig...@de.ibm.com>
* calls.c (store_unaligned_arguments_into_pseudos): Skip PARALLEL
arguments.
OK, so after a lot of worrying, I think this is OK. I kept thinking
this had to tie into the BLKmode return value braindamage that we have
to support on the PA. But that's not the case here. This is strictly
arguments.
If the argument is a PARALLEL, then yes, there's no sense in using
store_unaligned_arguments_into_pseudos, at least AFAICT, they are
handled by emit_group_move.
OK for the trunk, thanks for your patience,
Jeff
