Re: [PATCH] offline gcda profile processing tool

[Rong Xu <xur at google dot com>]

Here is the updated patch. The difference with patch set 3 is
(1) use the gcov-counter.def for scaling operation.
(2) fix wrong scaling function for time-profiler.

passed with bootstrap, profiledboostrap and SPEC2006.

Thanks,

-Rong

On Thu, May 1, 2014 at 3:37 PM, Rong Xu <xur@google.com> wrote:
> Hi, Honza,
>
> Please find the new patch in the attachment. This patch integrated
> your earlier suggestions. The noticeable changes are:
> (1) build specialized object for libgcov-driver.c and libgcov-merge.c
> and link into gcov-tool, rather including the source file.
> (2) split some gcov counter definition code to gcov-counter.def to
> avoid code duplication.
> (3) use a macro for gcov_read_counter(), so gcov-tool can use existing
> code in libgcov-merge.c with minimal change.
>
> This patch does not address the suggestion of creating a new directory
> for libgcov. I agree with you that's a much better
> and cleaner structure we should go for. We can do that in follow-up patches.
>
> I tested this patch with boostrap and profiledbootstrap. Other tests
> are on-going.
>
> Thanks,
>
> -Rong
>
> On Wed, Apr 16, 2014 at 8:34 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
>>> GCOT_TOOL needs to use this function to read the string in gcda file
>>> to memory to construct gcov_info objects.
>>> As you noticed, gcov runtime does not need this interface. But
>>> gcov-tool links with gcov runtime and it also uses the function.
>>> We could make it available in gcov_runtime, but that will slightly
>>> increase the memory footprint.
>>
>> Yep, it is not really pretty. I wrote bellow some plan how things may be
>> structured in more convenient way.  Any work in that direction would be welcome.
>>>
>>> The planned new functions for trunk version is: (1) overlap score b/w
>>> two profiles (2) better dumping (top hot objects/function/counters)
>>> and statistics.
>>> Once this basic version is in, we can start to add the new functionality.
>>
>> Sounds good. I assume the autoFDO does not go via gcov tool but rather uses
>> custom reader of profile data at GCC side?
>> I wonder, are there any recent overview papers/slides/text of design of AutoFDO
>> and other features to be merged?
>> I remember the talk from GCC Summit and I did read some of pre-LTO LIPO
>> sources & papers, but it would be nice to have somethin up to date.
>>>
>>> libgcov-util.o is built in gcc/ directory, rather in libgcc.
>>> It's directly linked to gcov-tool.
>>> So libgcov-util.o is built for HOST, not TARGET.
>>> With makefile changes, we can built HOST version of libgcov-driver.o
>>> and libgcov-merge.o.
>>> I also need to make some static functions/variables public.
>>
>> I suppose that can go with IN_GCOV_TOOL ifdef.
>>
>> So we currently have basic gcov io handling in gcc/gcov-io.c that is borrowed
>> by libgcc/libgcov* code.  We also will get libgcov-util.c in libgcc directory
>> that is actually borrowed by by gcc/gcov-tool.c code.
>>
>> We now have one runtime using STDIO for streaming and kernel has custom version
>> that goes via /proc interface (last time I looked).  We added some abstraction
>> into libgcov-interface that is the part that relies on STDIO, partly via gcov-io.c
>> code and now we have in-memory handling code in libgcov-util.
>>
>> I guess it would make most sentse to put all the gcov code into a new directory
>> (libgcov) and make it stand-alone library that can be configured
>> 1) for stdio based runtime as we do not
>> 2) for runtime missing the interface and relyin on user providing it
>> 3) for use within gcov file manipulation tools with reorg of
>> GCC/gcov/gcov-dump/gcov-tool to all use the same low-level interfaces.
>> In a longer term, I would like to make FDO profiling intstrumentation to happen
>> at linktime. For that I need to make the instrumentation code (minimal spaning
>> tree & friends) to work w/o cgraph that would ideally be done in a shared
>> implementation
>>> > Won't this get wrong answer when counters[0] is not the same?
>>> > I would expected the merging code to compare the counters first. Similarly for delta counter.
>>>
>>> These *_op functions are for scaling only. So there is only one
>>> profile involved (thus there is no comparison).
>>> The merge handles are in libgcov-merge.c which have the code to handle
>>> mismatched profile targets.
>>
>> I see, OK then.
>>> >
>>> > Adding correct rounding may actually make difference for Martin's startup time work.
>>>
>>> Do you mean to use something like in RDIV macro?
>>
>> Yes.
>>
>> Honza

Attachment: gcov_tool_patch_v4.txt
Description: Text document