On 03/11/16 12:05, Kyrill Tkachov wrote:
On 02/11/16 11:36, Eric Botcazou wrote:
I think you're right. I suppose the new condition should be:
#ifdef LOAD_EXTEND_OP
/* If this is a typical RISC machine, we only have to worry
about the way loads are extended. */
if (!WORD_REGISTER_OPERATIONS
|| ((LOAD_EXTEND_OP (inner_mode) == SIGN_EXTEND
? val_signbit_known_set_p (inner_mode, nonzero)
: LOAD_EXTEND_OP (inner_mode) != ZERO_EXTEND)
||
|| !MEM_P (SUBREG_REG (x))))
#endif
Agreed.
Would you prefer me to make this change or just revert the patch?
Go ahead and make the change, but please do a bit of comment massaging in the
process, for example:
#ifdef LOAD_EXTEND_OP
/* On many CISC machines, accessing an object in a wider mode
causes the high-order bits to become undefined. So they are
not known to be zero. */
if (!WORD_REGISTER_OPERATIONS
/* If this is a typical RISC machine, we only have to worry
about the way loads are extended. */
|| ((LOAD_EXTEND_OP (inner_mode) == SIGN_EXTEND
? val_signbit_known_set_p (inner_mode, nonzero)
: LOAD_EXTEND_OP (inner_mode) != ZERO_EXTEND)
|| !MEM_P (SUBREG_REG (x))))
#endif
{
if (GET_MODE_PRECISION (GET_MODE (x))
> GET_MODE_PRECISION (inner_mode))
nonzero |= (GET_MODE_MASK (GET_MODE (x))
& ~GET_MODE_MASK (inner_mode));
}
Thanks, here is the patch doing this.
Committing to trunk after bootstrap and testing on x86_64.
Sorry for the trouble,
Kyrill
With the following ChangeLog entry:
2016-11-03 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* rtlanal.c (nonzero_bits1): Fix WORD_REGISTER_OPERATIONS condition.
Move comments into more natural position.