Mainline bootstrap failure on ia64
Jim Wilson
wilson@specifixinc.com
Tue Apr 20 04:07:00 GMT 2004
Andreas Schwab wrote:
> Mainline does not bootstrap on ia64 because -frename-registers is now the
> default.
The problem is that the IA-64 port uses the common idiom of leaving off
the constraints for op0 of a call_value pattern, so that we can accept
any return rtx. In particular, we need to accept PARALLEL here which is
not an ordinary operand. However, the lack of constraints means we have
an output operand with no "=".
regrename handles conditional execution by identifying output operands
(i.e. operands with = constraints), and then changing them to in-out
operands if they occured inside a COND_EXEC. This fails when we have a
call_value pattern inside a COND_EXEC because the = constraint is
missing, so this operand does not get changed to an in-out operand.
This can be fixed by changing the call_value op0 constraints to "=X,X".
It looks a bit odd to use using X outside a match_scratch, but it
seems to do what I want, i.e. accept all operands, even things that
don't look like operands like PARALLEL.
Any other target that uses COND_EXEC and drops the constraints for op0
of a call_value pattern will run into the same problem. Perhaps we
should go through and fix all md files to use =X constraints here.
This patch works for the libstdc++-v3/src/wlocal-inst.cc testcase which
is the first failure I saw when doing an IA-64 bootstrap. I haven't
done any further testing yet.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: patch.ia64.regrename
URL: <https://gcc.gnu.org/pipermail/gcc/attachments/20040420/093bcb55/attachment.ksh>
More information about the Gcc
mailing list