On 12/03/13 15:46, Jakub Jelinek wrote:
Hi!
As described in the PR, the problem here is that during combine
i2 pattern is substituted into more than one place in i3 pattern,
unique_copy is 0 (but, even if it would be non-zero, it could happen
if the comparison was processed first before the normal set inside of the
parallel) and thus the same RTL is (temporarily) shared between two
locations. force_to_mode is first called with mask 0xdc36 (that is
actually find for both occurrences in the andhi_2 pattern) and later on
inside of the comparison again with mask 0x8000, and as it modifies
the IF_THEN_ELSE in place, it modifies also the other location (it is fine
if the comparison uses 0x8000 mask, but not in the other spot).
As in the end we fold it to a constant, we don't undo it and use incorrect
constant.
Fixed by making sure force_to_mode doesn't modify x in place.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk/4.8?
2013-12-03 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/58726
* combine.c (force_to_mode): Fix comment typo. Don't destructively
modify x for ROTATE, ROTATERT and IF_THEN_ELSE.
* gcc.c-torture/execute/pr58726.c: New test.
I'd worry there's other latent bugs of this nature and if we'd be better
off avoiding the temporary sharing. We have structure sharing rules for
a reason -- I'd hate to think of all the code that would need auditing
to ensure it was safe with this bogus sharing.
Any thoughts of how painful it'd be to avoid the sharing to start with?
jeff
