[LLVMdev] IC profiling infrastructure

[Betul Buyukkurt]

Thanks for all the replies.

> Xinliang David Li <davidxl at google.com> writes:
>> Justin, thanks for the reply.
>>
>> I would like to point out that value_kind is actually not really
>> stored in the raw profile data (nor do we intend to do so), other than
>> the minimal information about per value kind  NumOfSites info in
>> per-function profile header. Basically the data is organized per value
>> kind instead of stored intermixed. That is about it.
>>
>> More replies below.
>>
>> On Mon, Jun 29, 2015 at 11:53 AM, Justin Bogner <mail at justinbogner.com>
> wrote:
>>> I still have some issues with what we're doing with value_kind here.
> Let
>>> me summarize the arguments I've heard for including the value_kind in
>>> various places in this set of changes:
>>>
>>> 1. We would like to be able to enable multiple types of value profiling
>>>    at the same time.
>>>
>>> 1a. We would like to be able to selectively use only particular value
>>>     kinds from a particular profile.
>>>
>>> 1b. We would like to be able to combine profiles with different sets of
>>>     profiling options enabled.
>>>
>>
>> yes.
>>
>>> 2. We will need to know the value kind in order to parse the raw
> profile
>>>    data.
>>>
>>
>> Yes -- but we don't need to store value kind in order to do that -- we
>> simply need to organize the profile data more structurally.
>>
>>
>>> 3. We'd like to avoid needing to change the .profdata format when we
> add
>>>    new value kinds, since it needs to stay backwards compatible.
>>>
>>> 4. The frontend work is simpler if the value kind is explicitly
> encoded.
>>>
>>> There are a couple of these points that certainly need to be addressed,
>>> and some that are a little bit less clear.
>>>
>>> A. We do need to handle (1) in some way, but the issue can be stated a
>>>    little bit more generally. If there are optional parts of the
> profile
>>>    data, we definitely need to know which options are enabled.
>>>
>>> B. (1a) is only an issue if you believe (4). This might be true or it
> might
>>>    not, and I don't think we can make an informed decision until we're
>>>    actually using multiple value kinds. For now there's just a bunch of
>>>    code that's effectively useless - "if value kind is 2, do nothing".
> I'm
>>>    opposed to adding this kind of practically unreachable code "just in
>>>    case". It adds a kind of complexity and premature generality that,
> in my
>>>    experience, will have to be completely re-done when we get to
> actually
>>>    generalizing the code.
>>
>> In terms of value profiling code, there is no need for the check you
>> mentioned "if value kind is ..." at all, but 'attached' to the codegen
>> code for language construct of interests, for instance,
>>
>> ... CodeGenFunction::EmitCall(...)
>> {
>>
>>     if (enable_indirect_profile) {
>>            ....profileIndirectCall();     // does both instrumentation
>> and profile annotation depending on the context.
>>     }
>>    ...
>> }
>>
>> Organizing the profile data according value profile kind allows
>> profile data to be independently loaded and used:
>>
>> .. CodeGenPGO::assignRegionCounters (...) {
>>
>>    if (enable_indirect_profile) {
>>      loadValueProfile(VK_indirect_call);
>>    }
>>    if (enable_xxx_profile) {
>>     loadXXXProfile(VK_xxx);
>>   }
>>
>> }
>>
>> if  we end up with inter-mixing all value kind profile data, we will
>> have to load all value profile data into memory regardless of the
>> option specified.   It also makes the profile data uses really hard --
>> how do we easily match the value profile site with the profile data
>> counter index if profile kinds are selectively enabled?
>>
>> Besides, in the profile-use pass, each value kind's in-memory format
>> of profile data may be slightly different, so the loader will have to
>> do different things ..
>>
>>>
>>> C. (1b) is an interesting one. Is this actually a useful feature? If we
>>>    want to combine multiple profiles with different subsets of
> profiling
>>>    turned on, then we definitely need the information recorded
>>>    somewhere.
>>
>> I think it is a good flexibility to have.
>>
>>>
>>> D. I have issues with (2). If the different value kinds actually have
>>>    different structures and need to be read in / interpreted
>>>    differently, then writing the support for them now simply won't
>>>    work.
>>
>> We choose to have very compact form representing the raw value data
>> for all value kinds, but to support efficient use of value profile of
>> different kinds, the raw profile reader has to do something different
>> here. For instance, indirect call profile data needs the translation
>> of the raw address which other value kinds do not care.  I think this
>> is something we can do without.
>>
>>>    We're reading them in as if they're interpreted in the exact
>>>    same way as the indirect call profiling - if they aren't, then we
>>>    have to rewrite this code when we add a new one anyway. What's the
>>>    point of writing code that won't actually work when we try to use
> it?
>>
>> I don't think this would be a problem as adding a new kind requires
>> client/use side change too, otherwise we won't be able to use it.
>
> This is exactly my point. Why are we writing code *now* to read values
> that we have no way of using? This code *cannot* be tested, because it
> doesn't do anything. It adds a bunch of loops to iterate over values
> that *must* be zero, because if they aren't we don't actually know what
> they mean.

I disagree that the loops are doing nothing and the code is not testable.
First IR tests will come with path #4 which is the lowering. So it all is
going to be tested in a full-fledged form. Splitting was your suggestion.

> The compiler-rt patches I've seen so far have all treated each value
> kind as if it should be treated identically to indirect call target, and
> had extra complexity so that they could do this, but you've stated
> multiple times that this won't actually be correct when we use it. We
> should write the raw profile writer and reader code to do exactly what
> we need it to do today. We can and will change it when we want to add
> new functionality.

So I do not understand why having value kind hurts here. Just having
indirect call targets is the same as having a value_kind with size = 1.
Isn't it? We're also looking into profiling loop count histogram in the
near future. So, having a general value profiling foundation is extremely
useful. Otherwise, for each added value type the indexed reader format
will keep on changing. At least it's looking future proof in the sense
that new values up to three more are already being supported.

Knowing the value_kind is also necessary to be able to map the 64-bit
target addresses to function names. In the indexed format we either store
a String or a uint64_t value depending on the value_kind. I'm unable to
read from your arguments, how would you plan to store a name for one
specific value kind and a 64-bit value for another kind.

>>>
>>> E. There is value in dealing with (3) as long as we're confident that
> we
>>>    can get it right. If we're pretty confident that the value profile
>>>    data structure we're encoding in the .profdata file is going to work
>>>    for multiple value profile kinds, then it makes sense to allow
>>>    multiple of them and encode which kinds they are. David, is
>>>    "InstrProfValueRecord" in Betul's current patches going to work to
>>>    record the other types of value profiling information you're
> planning
>>>    on looking at?
>>
>> The raw profile format is certainly good for other kinds.
>
> I'm not talking about the raw profile format here. This point is about
> the stable, indexed format (.profdata, not .profraw). We have no need to
> keep the raw profile format backwards compatible, but we do need to keep
> readers for old indexed format files.
>
>> InstrProfValueRecord is mostly there (for instance overloading the
>> Name field for other purpose, but this is minor and can be done as
>> follow up when support of other kinds are added), so I am ok to leave
>> it as it is now.
>
> What do you mean? What do you expect this structure to look like for
> other value kinds? Will we need to change the serialized format for the
> indexed profile format?
>
> There are two possibilities. Either:
>
> 1. This serialized format is sufficient for other value kinds, so it is
>    valuable to encode it in the indexed profile format now, because this
>    will avoid needing to change that format later.
>
> Or,
>
> 2. This is likely not how other value kinds will need to be serialized,
>    so we're going to need to change the format later either way. In this
>    case, we should just solve the problem we have *today*, rather than
>    making changes we'll need to undo later.
>
> I think that given the uncertainty about what the future value kinds'
> formats will be, we should encode this in a way where it's obvious what
> the extra data is for (we'll specify a kind in the data), but without
> trying to accomodate future value kinds until we're ready to talk about
> them concretely.
>
> I'll go into more detail about this in my review for the indexed profile
> reader and writer patch, which I should be able to get to tomorrow.
>
>>> I'd also like to point out that the instrprof parts of LLVM try to be
>>> relatively frontend agnostic. We should strive to keep it such that the
>>> frontend is the piece that decides what a particular profiling kind
>>> means if possible, as long as the structure and way the data is laid
> out
>>> is suitable. This doesn't affect the current work very much, but it's a
>>> good idea to keep it in mind.
>>
>> yes. See the above -- FE code does not really need to care about the
> format.
>>
>>>
>>> So, with all of that said, I have a couple of suggestions on how to
> move
>>> forward.
>>>
>>> - I don't think we should store the kind in the raw profile yet. Let's
>>>   keep this simple, since it's easy to change and will be easier to get
>>>   right when we're implementing or experimenting with a second kind
>>>   (based on (D)).
>>>
>>
>> See my reply above -- we don't really store the value kind in raw
>> profile, but we need to organize the profile data according to the
>> kind in order to process (read) and use (profile-use) them
>> efficiently.
>
> The only reader of the raw profile is the tool that converts it to the
> indexed format, which is what profile-use consumes. The read will always
> read the entire file in a streaming fashion, and efficiently reading
> parts of the file is not important. We then write this to the indexed
> format, which needs to be efficiently readable.
>
>>
>>> - Assuming the answer to my question in (E) is "yes", let's go ahead
> and
>>>   include a value kind for the value records in the .profdata format.
>>
>> yes.
>>
>>>I
>>>   don't really like the current approach of an array of std::vectors,
>>>   because I think we can do this more simply. I'll bring that up on the
>>>   review thread.
>>>
>>> - In clang, we should error if we see value kinds we don't understand
>>>   yet.
>>
>> yes.
>>
>>>
>>> I think that this addresses the short term concerns without forcing us
>>> to write unreachable/untestable code that may or may not work for the
>>> longer term concerns. It minimally handles (A ) in that the data that
> is
>>> there implies which features are being used, puts off (B) without
> making
>>> it harder to deal with when the time comes, allows us to implement (C)
>>> without much trouble if we want it, avoids doing work that could cause
>>> problems with (D), and handles (E) since avoiding it would mean more
>>> work later.
>>>
>>> What do you think?
>>
>> See my reply above. For now, I think the things we need to organize
>> the value profile data in raw profile according to kind (not storing)
>> -- I think it will handle all the cases where you have concerns.
>>
>> thanks,
>>
>> David
>>
>>
>>
>>>
>>> Xinliang David Li <davidxl at google.com> writes:
>>>> Justin, do you have more concerns on keeping value_kind?
>>>>
>>>> If there is still disagreement, can we agree to move on with it for
>>>> now ?  After the initial version of the patches checked in, we can do
>>>> more serious testings with large apps and revisit this if there are
>>>> problems discovered.
>>>>
>>>> thanks,
>>>>
>>>> David
>>>>
>>>>
>>>>
>>>> On Mon, Jun 15, 2015 at 10:47 PM, Xinliang David Li
>>>> <davidxl at google.com> wrote:
>>>>>> But since the consumer is the frontend, and it knows which counts
> are
>>>>>> which, it knows the kind, no? I don't understand how storing the
> kind
>>>>>> info helps or hurts - it's just redundant.
>>>>>
>>>>> Yes, the frontend consumer knows, but the raw reader does not. It is
>>>>> natural to do those processing when processing the raw profile data
>>>>> (for instance when function pointer to name mapping still exists).
>>>>>
>>>>>>
>>>>>>>>
>>>>>>>> The other thing we should do is store which profiling options are
>>>>>>>> enabled, in both formats. When we specify profile-instr-use it'll
> be
>>>>>>>> less error prone if we can detect whether or not this includes
> indirect
>>>>>>>> call profiling without checking other options, and it's probably a
> good
>>>>>>>> idea for "llvm-profdata merge" to disallow merging profiles that
> are
>>>>>>>> gathering different sets of data. A bitfield seems suitable for
> this.
>>>>>>>
>>>>>>> For value profiling, allowing profile-gen and profile-use passes
> using
>>>>>>> different options is a useful feature. Consider the following
>>>>>>> scenarios:
>>>>>>> 1) collect the same profile data once with all kinds of value
> profile
>>>>>>>    data collected. The exact same profile data can be used in
> performance
>>>>>>>    experiments with different kinds of value profiling
> enabled/disabled
>>>>>>
>>>>>> This use case is pretty easy to accomodate for with the model where
> the
>>>>>> profile stores its type. We could autodetect the type to decide how
> to
>>>>>> interpret the profile, and if more specific profile flags are set
> simply
>>>>>> ignore some of the data. The thing that the file format storing the
> type
>>>>>> gives us is that it's harder to misinterpret the data by supplying
> the
>>>>>> mismatching flags between reading and writing.
>>>>>
>>>>> Yes, this is achievable with some extra effort -- for instance by
>>>>> collecting the profile sites of all kinds regardless of the flags.
>>>>> After profiling matching, simply drop selected sites according to the
>>>>> flags.
>>>>>
>>>>> This adds complexity of the code. With value kinds encoded in profile
>>>>> data, the handling is much simpler -- each value profiler only needs
>>>>> to deal with its own kind.
>>>>>
>>>>>>
>>>>>>> 2) work around profile-use/value transformation bug selectively for
>>>>>>>    some file without the need to change anything in instrumentation
> pass
>>>>>>
>>>>>> I'm not sure I understand what you mean here.
>>>>>
>>>>> Similar to the above, but for correctness.
>>>>>
>>>>>>
>>>>>>> Besides, with the latest patch, the value_kind is not recorded in
> each
>>>>>>> profile value thus the overhead is minimized.
>>>>>>
>>>>>> The overhead is low, sure, but the code complexity of dealing with
> the
>>>>>> multiple kinds in this way is quite real.
>>>>>
>>>>> I actually believe having value-kind can help reduce code complexity
>>>>> instead of the other way around. What is the extra complexity?
>>>>>
>>>>> thanks,
>>>>>
>>>>> David
>>>>>
>>>>>
>>>>>>Since I'm not convinced we're
>>>>>> getting real benefit from storing the kind, I don't believe this
>>>>>> trade-off is worth it.
>>>>>>
>>>>>>> thanks,
>>>>>>>
>>>>>>> David
>