Hi Martin, On Wed, Aug 6, 2014 at 4:02 PM, Martin Jambor <mjam...@suse.cz> wrote: > Hi, > > On Wed, Jul 30, 2014 at 06:56:05PM +0200, Martin Jambor wrote: >> Hi, >> >> IPA-CP can wreck havoc to transactional memory support as described in >> the summary of the PR in bugzilla. It seems the cause is that IPA-CP >> clones of nodes created by trans-mem do not have their tm_clone flag >> set. For release branches we have decided to simply disable IPA-CP of >> trans-mem clones but for trunk we'd like to avoid this. I am not 100% >> sure that just copying the flag is OK but it seems that it works for >> the provided testcase and nobody from the trans-mem people has >> commented in bugzilla for over a month. So I suggest we commit this >> patch and wait and see if something breaks. Hopefully nothing will. >> > > Honza has approved the patch in person and so I have committed it as > revision 213666 after re-testing. Hopefully it does not break > anything, if it does then read the paragraph above and remember it is > not really my fault :-)
Thanks for fixing this! Copying the flag seems to me the good solution. tm_clone flag indicates that this is an instrumented function. It is used for different part in trans-mem but particularly to find tm region to instrument. However, I am wondering if we should not call record_tm_clone_pair if it duplicates a tm_clone. libitm needs to find the clone of particular function. So the question is: Can the compiler replace a function pointer by its corresponding constprop function pointer in some cases? -- Patrick Marlier
