[Google/4_8] LIPO COMDAT profile fixups

[Teresa Johnson]

On Mon, May 19, 2014 at 11:51 PM, Xinliang David Li <davidxl@google.com> wrote:
> Why duplicating the merger functions in dyn-ipa.c? Should those in
> libgcov-merge.c be reused?

The merger functions in libgcov-merge.c use a macro to either read the
counter to merge from a buffer in memory (when IN_GCOV_TOOL) or from
the gcda file otherwise. In this case I need to merge from another
array, and it needs to do so both IN_GCOV_TOOL and otherwise. So to
reuse the merger functions I would have had to get the counter value
from a new argument, guarded by a conditional. I didn't want to change
the interface of those existing merge functions, or more importantly,
make them less efficient with the added condition since they are
invoked frequently during gcda file merging.

>
> The refactoring of gcov_exit_write_gcda should probably be done in a
> separate patch -- preferably submitted to trunk too.

The refactoring is necessary for this patch since I need to invoke the
outlined functionality in gcov_dump_module_info as well.

I could submit that to trunk, but it has to go in to the google
branches either with or preceeding this patch.

Thanks,
Teresa

>
> David
>
> On Mon, May 19, 2014 at 10:08 PM, Teresa Johnson <tejohnson@google.com> wrote:
>> Ping.
>> Teresa
>>
>> On Wed, May 14, 2014 at 4:39 PM, Teresa Johnson <tejohnson@google.com> wrote:
>>> This patch applies profile fixups to COMDATs on the dyn ipa callgraph
>>> at the end of LIPO module grouping (either in the profile gen run or
>>> in gcov-tool). This is to address issues with missing profiles in the
>>> out-of-line COMDAT copies not selected by the linker, and indirect
>>> call profiles that target a module not included in the module group.
>>> By default, both fixups are enabled, but can be controlled by a
>>> profile-gen parameter, an environment variable, and a gcov-tool
>>> option.
>>>
>>> The fixups assume that functions with the same lineno and cfg checksum
>>> are copies of the same COMDAT. This is made more likely by ensuring
>>> that in LIPO mode we include the full mangled name in the
>>> lineno_checksum.
>>>
>>> For the counter fixup, we merge all non-zero profiles with matching
>>> checksums and copy the merged profile into copies with all-zero
>>> profiles.
>>>
>>> For the indirect counter fixup, if an indirect call profile target is
>>> not in the module group, we look for a matching checksum copy in the
>>> primary module and if exactly one is found we change the target to
>>> that.
>>>
>>> If any fixups are applied, the gcda files are rewritten after module grouping.
>>>
>>> This also required a couple of other changes to the optimizer. During
>>> cgraph node resolution, we were arbitrarily selecting a copy that had
>>> non-zero profiles. Now there are many more to choose from, so we will
>>> prefer the primary module copy if it is non-zero. Also, during cloning
>>> for inlining, we only want to update the profile on the callee node if
>>> we are inlining into the resolved node caller node. We were already
>>> doing this for AutoFDO, and need to do this here now that many node
>>> copies have the same profile.
>>>
>>> Patch attached. Tested with regression tests and internal benchmarks.
>>> Ok for google branches?
>>>
>>> Thanks,
>>> Teresa
>>>
>>>
>>>
>>> --
>>> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413
>>
>>
>>
>> --
>> Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413



-- 
Teresa Johnson | Software Engineer | tejohnson@google.com | 408-460-2413