Alexey Varlamov said the following on 10.12.2007 9:33:
IMO this fix is a must for M4:
1) clearly this is a regression and should be fixed within the same
milestone as we agreed;
2) no point to distribute completely broken component in the milestone
build, so otherwise we'd need to drop the gc_cc from distribution.
As a side note, we unjustly left the gc_cc overboard BTI testing. IMO
we should include additional mode to DRLVM suites which support
multi-mode runs (smoke tests and EHWA-api). The only question is do it
now or let M4 finish...
I believe it would not impact testing time too much, <20% for those
suites (AFAIK they currently run 5 modes: jet,jit,opt,srv,int). Name
it "gc41" ?
+1 to include gc_cc into regular testing. I think inclusion may wait
until after M4 since according to Ilya there are still some bugs
remaining when using gc_cc.
2007/12/9, Ilya Berezhniuk <[EMAIL PROTECTED]>:
Thanks Xiao Feng.
But 24h later will be Monday, so '2 committers approval' is required
to check in the patches.
Committers,
please support the patches if no objections :)
Thanks,
Ilya.
2007/12/9, Xiao-Feng Li <[EMAIL PROTECTED]>:
I'd like to have GC_CC in M4. If no objections in 24 hours, I will
commit the patch. Thanks. -xiaofeng
On Dec 9, 2007 10:11 AM, Ilya Berezhniuk <[EMAIL PROTECTED]> wrote:
Hi All,
Now we have GC_CC inoperable on most platforms - it works on
Windows/x86/release only.
I've checked M3 build, GC_CC works fine (although hangs from time to time).
I suggest committing HARMONY-5247 and HARMONY-5262 patches to fix most
of current GC_CC problems.
The problems are the following:
1) After r583223 threading commit GC_CC does not work in debug mode
because of multiple assertions.
2) After r587472 commit (symbols exporting policy) GC_CC was broken on
Linux because several symbols required by GC_CC were not added to EXP
files.
3) After r599482 commit (uncompressed references) GC_CC works
incorrectly on x86_64 because of changed object layout for 64bit
platforms.
4) GC_CC often hangs in shutdown. HARMONY-5136 has fixed the problem
for GC_GEN, but for GC_CC the problem still exists because of
different GC/Finalizer threads behavior.
HARMONY-5247 patch finally fixes the problem with shutdown deadlock on
ThreadGroup.lock monitor.
HARMONY-5262 patch fixes GC_CC problems 1) - 3).
With both these patches all 'build test' tests are passed on
Windows/Linux x86 and on Linux/x86_64 in both debug and release modes.
On 4-core Windows/x86_64 I've got 4 intermittent failures on debug
build (see details in HARMONY-5262), but these failures are not
reproducible on release build, and they also do not appear on 2-core
machine.
Looks like these failures are caused by some remaining threading
problems in GC_CC.
Since HARMONY-5262 affects only GC_CC, all tests are passed with
GC_GEN on all 4 tested platforms.
I think it's valuable to make GC_CC working, if we are planning to
include it into M4 builds.
Thanks,
Ilya.
--
http://xiao-feng.blogspot.com
