jdk16 and -fno-tree-vrp

Björn König bkoenig at alpha-tierchen.de
Mon Nov 12 12:05:47 PST 2007


Alexander Kabaev wrote:

> 1. jdk15 is missing from the list of jdk's you tried to build, and that
> was the primary tree-vrp victim;

Ok, I did that too right now and it builds without problems either. By the
way, I used diablo-jdk15 for bootstrapping.

> 2. You did not mention architecture you did your experiments on.

i386. I plan to build the ports on amd64 too.

> This particular optimization step was implicated in a number if
> miscompiles reported by several developers and just taking it out of
> the picture takes care of majority. This is right decision too given
> where we are in the release cycle and unclear position of the project
> towards GPL3.

I understand the severity of this decision, but I'm a little bit in doubt.

My first point is that most of these reported bugs happend during the
pre-beta stage where -current was still not settled very well. From
experience we know that numerous users start to test it during the BETA or
even RC phase. Therefore I think that there were too less testers and too
less bug reports.

Furthermore I assume that somebody just tried -fno-tree-vrp, hit or miss,
it seem to work more or less coincidentally and I'm not sure if this is a
good solution. I think this because -fno-tree-vrp breaks the FreeBSD/arm
world which is definitely a GCC bug. It is strange that it works fine
again if you specify -fno-cse-follow-jumps or -fno-schedule-insns for
example. I know that that FreeBSD/arm has not the same importance as i386
or other tier 1 platforms. Even the jdk ports are more important for the
upcoming release. It is just an example to emphasise the quality of this
bug. So in other words you closed one box of pandora and opened another
one.

Last but not least I found a small change the the GCC repository which
unbreaks the ARM world compiled with -fno-tree-vrp and I'm curious whether
the same change affects the jdk ports. So I build them first with the
assumption that they will fail, but they didn't. That's also why I ask for
a reproducible scenario.

Björn




More information about the freebsd-java mailing list