https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70222
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So, the bug is apparently in simplify_shift_const_1 (code=LSHIFTRT,
result_mode=SImode, varop=0x7ffff184de70, orig_count=31)
where varop is:
(subreg:SI (lshiftrt:DI (const_int -1 [0xffffffffffffffff])
(subreg:QI (reg:SI 100) 0)) 0)
We optimize the inner shift (mode == DImode) into
(lshiftrt:DI (const_int 8589934591 [0x1ffffffff])
(subreg:QI (reg:SI 100) 0)) 0)
which is fine (assuming we do the final masking), and turn shift count into 0.
shift_mode is still SImode, because nonzero_bits doesn't say the upper bits are
0, and we don't have OUTER_OP. For this to really work, we need to mask off
the bits though.
We have:
/* If we were doing an LSHIFTRT in a wider mode than it was originally,
turn off all the bits that the shift would have turned off. */
if (orig_code == LSHIFTRT && result_mode != shift_mode)
x = simplify_and_const_int (NULL_RTX, shift_mode, x,
GET_MODE_MASK (result_mode) >> orig_count);
that does this, unfortunately it doesn't trigger here, because shift_mode ==
result_mode == SImode.
So, one possibility is to do
if (orig_code == LSHIFTRT
&& (result_mode != shift_mode
|| (result_mode != mode && count == 0)))
(and maybe "&& orig_count != 0" too?), to say that we need to mask even if the
shift in shift_mode actually isn't performed at all (or shall it be orig_count
!= count ?).
Or we'd need to force outer_op in this case somewhere, but not sure what all
cases would need to do that.
