[PATCH] D29512: [PGO] Directory name stripping in global identifier for static functions

Tue Feb 7 18:55:23 PST 2017

On Mon, Feb 6, 2017 at 10:27 AM, Rong Xu <xur at google.com> wrote:

>
>
> On Sun, Feb 5, 2017 at 12:11 AM, Xinliang David Li <davidxl at google.com>
> wrote:
>
>>
>>
>> On Fri, Feb 3, 2017 at 11:28 PM, Sean Silva <chisophugis at gmail.com>
>> wrote:
>>
>>>
>>>
>>> On Fri, Feb 3, 2017 at 10:18 PM, Xinliang David Li <davidxl at google.com>
>>> wrote:
>>>
>>>>
>>>>
>>>> On Fri, Feb 3, 2017 at 8:31 PM, Sean Silva via Phabricator <
>>>> reviews at reviews.llvm.org> wrote:
>>>>
>>>>> silvas added a comment.
>>>>>
>>>>> This change does two things (as you mention in the description):
>>>>>
>>>>> 1. Adding -static-func-strip-dirname-prefix which provides a way to
>>>>> have more control when `-static-func-full-module-prefix=true` is
>>>>> specified.
>>>>>
>>>>
>>>> This is actually a more general form of -static-func-full-mdoule-prefi
>>>> x.
>>>>
>>>>
>>>>> 2. Changing the default of -static-func-full-module-prefix to true.
>>>>>
>>>>> IIRC, -static-func-full-module-prefix defaults to false because it
>>>>> caused issues when set to true (in fact, it was introduced to avoid these
>>>>> issues). The default value of -static-func-strip-dirname-prefix
>>>>> introduced in this patch (i.e. 0) is effectively a no-op; so ignore 1. for
>>>>> now. This means that the net effect of this patch is that compilation will,
>>>>> by default, have a regression on the issue fixed by r275193 /
>>>>> http://reviews.llvm.org/D22028, which is not a good idea. I think
>>>>> that the default behavior (which is user-visible) should not be changed in
>>>>> this patch.
>>>>>
>>>>
>>>> I disagree. The original default behavior was to preserve the full path
>>>> which was also user visible :)
>>>>
>>>
>>> And yet we found a compelling-enough use case to change it.
>>>
>>
>> It was probably better to introduce the option but not flipping the
>> default the first time. The stripping-path-fully mode had not been widely
>> tested at that time.
>>
>> I am a little curious about the use case for D22028. The pgo name of
>> static function is only affected by source module path. Why would that be
>> different for pgo-gen/use builds? In most common setup I saw, the source
>> paths should remain the same.
>>
>>
>>> We may need to revisit that decision, but clearly the current default is
>>> intentional and part of changing away from that is explaining why we no
>>> longer care about that use case (or care about it less than some other
>>> thing).
>>>
>>
>> We care about all use cases, which is why the more general form of option
>> is introduced -- it makes sure the use case in D22028 can also be handled
>> but more safely (stripping all prefix will bound to cause problems).
>>
>>
>>>
>>>
>>>>
>>>> The whole rationale for changing the current default is that it is
>>>> generally not safe -- mainly problem #1 because of counter variables for
>>>> static functions can not guaranteed to be unique when full path is stripped.
>>>>
>>>
>>> Can the counter variables be static to match the static nature of the
>>> functions they describe? (there would still be collisions when indexing the
>>> profile data though; the function CFG hash could be included in the "name"
>>> to avoid this)
>>>
>>
>> For simple functions, cfg hash collision is also very likely, so the
>> first line of defense is always the name key.
>>
>>
>>>
>>>
>>>> The ThinLTO issue is secondary (probably irrelevant here because of
>>>> other bugs).
>>>>
>>>
>>>>
>>>> The issue addressed in D22028 is actually not common -- the source
>>>> module paths should generally match in profile-gen and profile-use phases,
>>>> so using internal option for that use case seems more reasonable to me.
>>>>
>>>
>>> This is subjective, but I think it is quite reasonable to assume that
>>> each build will use a different output directory. Hence any build that
>>> generates .cpp files into the output directory (which seems reasonable too)
>>> is susceptible.
>>>
>>
>> so the case is for generated source files? Should they be accessed with
>> relative paths?
>>
>>
>>>
>>> Overall, requiring a user to use a compiler-internal option for
>>> something that seems to happen in practice (e.g. back when I was a
>>> PlayStation we actually ran into it and spent time fixing it) is a pretty
>>> poor experience. I think we should aim to do better (though we might settle
>>> for less if that proves challenging).
>>>
>>
>> I agree in general. However I think it is reasonable for a user to use an
>> internal option for corner use cases.  Another choice is to introduce an
>> external option for this which user can rely on.
>>
>>>
>>>
>>>>
>>>>
>>>>
>>>>>
>>>>> Overall, it sounds like this approach of relying on users to tweak
>>>>> internal compiler options (-mllvm) to get correct behavior in their
>>>>> environment is not the kind of user experience we want to deliver (or the
>>>>> kind of implementation that we want to maintain). IIRC, when we added
>>>>> -static-func-full-module-prefix, it was with the understanding that
>>>>> it was a simple hack for working around the larger issue of relying on the
>>>>> module name which we knew was not very robust. The further addition of the
>>>>> "InLTO" complicates things even further. It seems like a code smell that we
>>>>> do not have a Single Point Of Truth.
>>>>>
>>>>> I proposed a solution at one point https://groups.google.com/d/ms
>>>>> g/llvm-dev/s_VZbFTWbVs/d0b4Zh80CgAJ though it may no longer be
>>>>> applicable. It seems like ThinLTO already has to solve a problem of finding
>>>>> unique identifiers for all functions (even static), so we may want to
>>>>> piggy-back on that mechanism (this is just a high-level thought; haven't
>>>>> looked into the details).
>>>>>
>>>>>
>>>> For LTO/ThinLTO,  we solved the issue by using meta data which uses
>>>> getPGOFuncName as singe source of truth.
>>>>
>>>
>>>>
>>>>
>>>>> So:
>>>>>
>>>>> - I specifically object to changing user-visible defaults in this
>>>>> patch. Those changes should be isolated, and I don't think we have
>>>>> justification to change those defaults anyway.
>>>>>
>>>>
>>>> See my reply about the safety issue of keeping the current default.
>>>>
>>>>
>>>>
>>>>> - I'm slightly opposed to adding the -static-func-strip-dirname-prefix
>>>>> flag, since it seems like a workaround (among others that have already
>>>>> piled up) for a more fundamental issue. This is a frog-in-boiling-water
>>>>> situation; if solving the fundamental issue would be a huge amount of work,
>>>>> then adding the new flag is probably fine for now, but we need to keep in
>>>>> mind the larger situation. IIUC, defaulting `-static-func-strip-dirname-prefix=-1`
>>>>> would emulate the current default behavior, so
>>>>> -static-func-full-module-prefix could just be removed in the same
>>>>> patch.
>>>>>
>>>>
>>>> The current -static-func-full-module-path=false is simply a special
>>>> case of the new option. For users who rely on this option may hit the
>>>> correctness issue, they won't have any fallback without the new option.
>>>>
>>>>
>>>>> - I would encourage brainstorming/discussion of alternative solutions
>>>>> that solve the fundamental problem (which seems to be more about having a
>>>>> stable globally unique identifier than being specifically about
>>>>> preserving/mangling the "name" per se).
>>>>>
>>>>
>>>> The problem itself is simple: handle name conflicts between
>>>>
>>>> /a/b/c/foo.c:static_func
>>>> /e/f/g/foo.c:static_func
>>>>
>>>> Path info is a natural choice. Note that FE instrumentation also uses
>>>> module path to uniquely identify static_func as well.
>>>>
>>>
>>> Yes, the problem was inherited from FE instrumentation. I remember that
>>> when I explained to Justin the issue, he said that it was clearly buggy and
>>> not intentional (an oversight when implementing FEPGO).
>>>
>>
>> It works very well in practice -- though it is not guaranteed to be 100%
>> free of conflict. I won't label it as buggy.
>>
>>
>>>
>>> It seems that the fundamental issue is coming up with a unique
>>> identifier for the current TU that is stable across compiler invocations.
>>> How do other compilers handle this?
>>>
>>> GCC does not suffer from the problem because it does not dump profile
>> into one file but one profile file per module. The profile data file tree
>> structure mirrors the build output file structure so there will be problem
>> if profile-gen and use do not share the same structure. Runtime options are
>> provided to strip prefixes from output directories. Compiler time option is
>> also provided to relocate profile data (e.g. pointing to different root).
>>
>>
>>
>>> For example, path names are not enough. E.g. a user may build
>>> /a/b/c/foo.c with two different sets of compiler options, yet static
>>> functions of the same name must still be treated as separate. A file like:
>>>
>>> foo.c:
>>>
>>> static void PreprocessHelper(/* something */) {
>>> #ifdef USE_AVX
>>>   // something
>>> #else
>>>   // something else
>>> #endif
>>> }
>>>
>>> #ifdef USE_AVX
>>> void foo_avx(/* something */) {
>>> #else
>>> void foo_noavx(/* something */) {
>>> #endif
>>>   PreprocessHelper(/* something */);
>>>   // something
>>> }
>>>
>>>
>>> IIRC, one option (suggested by pcc if I remember correctly) is to use a
>>> hash of the TU's exported symbols (or something like that) to uniquely
>>> identify the TU.
>>>
>>
>> Should be limited to user defined public symbols.
>>
>>
>>> That seems more robust than a path name.
>>>
>>
>> I believe so.
>>
>> Content based ID has its advantage but has disadvantages too. For
>> instance more expensive to compute, less readable names. Using path based
>> naming, we can immediately identify where the static function is defined.
>> Perhaps we can use base name plus content hash.
>>
>> Teresa, Rong, do you see a situation when  module ID needs to be
>> identified but it is difficult or  too expensive to access the module's
>> content?
>>
>
> Another disadvantage is this is less tolerable to source changes: if the
> use adds another exported symbol in the use-compilation (which has nothing
> to do with another of the existing functions), the suddenly voids all the
> static function's profile.  This is not happening in current path name
> based scheme.
>

You are arguing that this is a correctness issue, and I gave an example
where the path based scheme is incorrect. Conservatively invalidating
profile info seems preferable, no?

-- Sean Silva

>
>
>
>>
>> David
>>
>>
>>> -- Sean Silva
>>>
>>>
>>>>
>>>> thanks,
>>>>
>>>> David
>>>>
>>>>
>>>>
>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> https://reviews.llvm.org/D29512
>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170207/3bdb3a4f/attachment.html>