This is the mail archive of the
mailing list for the GCC project.
Re: define_constraints patch, re-revised
On 10/2/06, Zack Weinberg <email@example.com> wrote:
Oh ho, I know what's going on now. There are two problems. First, I
was mistaken in thinking that the boilerplate code got unconditionally
emitted; it's cut out if we don't think it's needed. Second, there is
an optimization in add_constraint ("Remove the redundant match_code
from the expression") which is correct when IJKLMNOP constraints are
called from the legacy interface (CONST_OK_FOR_CONSTRAINT/LETTER) but
not when the satisfies_constraint_* functions / satisfies_constraint_p
are called directly. I will try to find time to patch this in the
next week or so, but I can't make any promises.
Here is a patch which I believe will fix this problem and a number of
related problems. Assuming this message gets through - I keep getting
"quota exceeded" bounces - would you please test it on your
architecture, Rask? It has passed an i686-linux bootstrap, and manual
inspection of the relevant generated files suggests the problem is
gone, but that target is not terribly interesting.
* genpreds.c (add_constraint): Do not remove anything from the
(write_tm_constrs_h): Determine whether op will be unused, and
tag it with
ARG_UNUSED if so.
(write_insn_const_int_ok_for_constraint): Avoid writing out
the redundant (and
impossible to evaluate in this context) test for a CONST_INT
that will be in the predicate