[PATCH 1/2] New pass to partition single function into multiple sections (take 2)

Revital1 Eres ERES@il.ibm.com
Sun Oct 19 17:46:00 GMT 2008


Hello,

> > > I regtest with -O2 (which includes several testscases for hot/cold
> > > partitioning). I will double check they are OK. Should I run other
tests
> > > to validate that?
> > >
> > > Revital
> >
> > Jan is the expert on profiling.  I just wanted to make sure the
hot/cold
> > profiling still worked with the changes.
>
> Jan - is there any other tests you think I should do to validate the
> hot/cold profiling?
>
> Trevor - does the SPU part is OK with you?

Jan recently pointed out that the cfg together with profiling information
can be retrieved after free_cfg pass (I was under the impression
it can not be done); so the option of rescheduling this pass after
pass_machine_reorg was raised again (using compute_bb_for_insn () to
retrieve the cfg).

After spending some time investigating this option it seems that choosing
this road might lead to other compilation.  For the SPU back-end hints
should be inserted for the new branches created between sections;
inserting them after machine reorg pass might mess up the alignment and
the careful insertion of nops.  Leaving the pass as is (scheduled
before free_cfg) and let the machine reorg pass work once after the
partitions were created seems to be the cleanest solution.

Attached is the patch adjusted to current mainline. (no major changes
were made)
It passes bootstrap and regtest on powerpc64 and x86_64. On SPU it
passes regtest (c, c++, and fortran).

Jan, Trevor - does this patch OK for mainline?

Thanks,
Revital


(See attached file: Change_log_8_9.txt)(See attached file:
patch_testsuite_7_9.txt)(See attached file: patch_partition_12_10.txt)
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: Change_log_8_9.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20081019/c8115ed1/attachment.txt>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: patch_testsuite_7_9.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20081019/c8115ed1/attachment-0001.txt>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: patch_partition_12_10.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20081019/c8115ed1/attachment-0002.txt>


More information about the Gcc-patches mailing list