[PATCH] D20348: IR: Introduce local_unnamed_addr attribute.

Tue May 31 18:11:14 PDT 2016

> On May 31, 2016, at 6:00 PM, Chandler Carruth <chandlerc at google.com> wrote:
> 
> On Fri, May 27, 2016 at 8:43 AM Mehdi Amini via llvm-commits <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>> On May 18, 2016, at 11:16 AM, Peter Collingbourne via llvm-commits <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>> 
>> 
>> 
>> On Tue, May 17, 2016 at 6:46 PM, Chandler Carruth <chandlerc at gmail.com <mailto:chandlerc at gmail.com>> wrote:
>> On Tue, May 17, 2016 at 6:40 PM Peter Collingbourne via llvm-commits <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>> On Tue, May 17, 2016 at 6:07 PM, Chandler Carruth <chandlerc at gmail.com <mailto:chandlerc at gmail.com>> wrote:
>> Thanks for the detailed write-up, and sorry to Rafael and Mehdi that it's on a new thread. =/
>> 
>> On Tue, May 17, 2016 at 5:59 PM Peter Collingbourne via llvm-commits <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>> pcc created this revision.
>> pcc added reviewers: rafael, joker.eph, chandlerc, majnemer.
>> pcc added a subscriber: llvm-commits.
>> Herald added a reviewer: tstellarAMD.
>> Herald added subscribers: jfb, mzolotukhin, joker.eph, arsenm.
>> 
>> If a local_unnamed_addr attribute is attached to a global, the address
>> is known to be insignificant within the module. It is distinct from the
>> existing unnamed_addr attribute in that it only describes a local property
>> of the module rather than a global property of the symbol.
>> 
>> This attribute is intended to be used by the code generator and LTO to allow
>> the linker to decide whether the global needs to be in the symbol table. It is
>> possible to exclude a global from the symbol table if three things are true:
>> - This attribute is present on every instance of the global (which means that
>>   the normal rule that the global must have a unique address can be broken without
>>   being observable by the program by performing comparisons against the global's
>>   address)
>> - The global has linkonce_odr linkage (which means that each linkage unit must have
>>   its own copy of the global if it requires one, and the copy in each linkage unit
>>   must be the same)
>> - It is a constant or a function (which means that the program cannot observe that
>>   the unique-address rule has been broken by writing to the global)
>> 
>> Although this attribute could in principle be computed from the module
>> contents, LTO clients (i.e. linkers) will normally need to be able to compute
>> this property as part of symbol resolution, and it would be inefficient to
>> materialize every module just to compute it.
>> 
>> Cool, this last part is really key.
>> 
>> 
>> My real problem with adding this as a normal attribute is that I'm not sure what it really means. Is it just a "cache" of some local analysis? Do we expect things to invalidate it if they make the address significant within a module? Is this something that would be "blessed" by some frontends?
>> 
>> I feel like, from your description, this really is just intended to solve the problem of materializing all of the module. It would seem that for that purpose something more akin to the "summary" information used by ThinLTO would be a better tool than an attribute which has to have a semantic contract for the IR.
>> 
>> What do you think?
>> 
>> To a certain extent this is a summary of the module contents, and in most cases I'd expect the property to simply summarize the module.
>> 
>> However, it is also a property that should be preserved if a pass introduces an address comparison. Modulo bugs, the new comparison should be "benign" -- if the original program could not observe the address, the optimized program shouldn't be able to observe it either.
>> 
>> One example of this would be a comparison of a vptr against a vtable or function address for speculative devirtualization, or in general any form of PGO that relies on global addresses. The introduction of such a comparison doesn't invalidate the unnamed_addr property (which we do currently apply to vtables), as the program would still have the same semantics if we, say, merged two identical vtables. The same applies to the local_unnamed_addr property.
>> 
>>  Regarding frontends, yes, I'd expect that if a frontend knows that all address comparisons within a module are benign, it could apply this property.
>> 
>> OK, all of this argues that we *can* define this as a semantic attribute, but doesn't really speak to why we *should*.
> 
> 
> As Peter mentioned above, some transformations can introduce constructs that would prevent from inferring the attribute, which means that when writing out bitcode we won't be able to generate it.  Because of that, having it only in the bitcode (like the summary) makes it more "fragile" (round-tripping may be broken).
> 
> It seems to me that if we don't make it first class in the IR, we will still want to have an analysis (or metadata) populated from the summary data that can be used to preverse this "property".
> 
> An analysis to allow easy access makes perfect sense to me, and even something more akin to metadata might make sense.

I don't see the fundamental difference between a metadata and an attribute when you start to attach it some semantic, it looks like an implementation detail to me. Maybe the difference is that we should always be able to drop a metadata without correctness issue? But it seems to me that this would apply to multiple existing attributes as well.

> 
> I'll try to explain why an IR attribute seems strange to me: we have to remember for all time to update it.
> 
> I think that IR attributes which only *reflect* or "cache" the state of the program itself make every transformation which could possibly invalidate them fragile. We have to continually remember the set of attributes to go and invalidate. Attributes seem much more important when they *promise* some state that may not be (re-)computable, and thus it is *necessary* to have a fundamental semantic bit to indicate that the property must hold. Even better when the properties themselves are inherently defined in a way that isn't invalidated by transformations.

Isn't it exactly the case here? 
Transformations will not invalidate it "conceptually" (the attribute would always be valid and we don't need to care about invalidating it), but on the other side transformations can turn the code into a form where we can't deduce the attribute anymore (i.e. what you described IIUC as 'we "promise"  that may not be (re-)computable').
It is possible that the set of transformations that prevents from inferring it too small to care about loosing this information?

-- 
Mehdi

> 
> "readonly" is IMO a great example. Sometimes we compute it, but there are plenty of times where we can't and still want the knowledge. And to mitigate the cost this still imposes, the set of transformations that cause a function to *start* to mutate external state are very, very rare. To justify the cost, the value it provides compared to an analysis is *huge*.
> 
> When we can trivially compute the result (or the value of representing a-priori information is so small that we're willing to forego it), we should rely on things that "self invalidate" like analysis passes.
> 
> All this said, I do agree that this makes things like round tripping somewhat more fragile. It's just that I think that is a better tradeoff than growing the set of possible semantic bits that we have to be aware of while writing every transformation.
> 
> -Chandler
>  
> 
> -- 
> Mehdi
> 
> 
>> 
>> Adding yet another semantic attribute is a really invasive change, and it feels like there should be a smaller / simpler mechanism to achieve the goals you have here. The whole point of having summary information available it bitcode was to allow LTO-like applications to read a small header to answer specific questions like the one motivating this patch.
>> 
>> I'd like to understand why a summary approach isn't the correct approach here before we extend the IR to represent a new concept, especially one with extremely close overlap and subtle distinctions from an existing concept.
>> 
>> We could make this a summary or an analysis, but that wouldn't really change anything as I see it. Whatever representation we use would need to be:
>> - stored in the bitcode
>> - available at LTO and (as Pete pointed out) codegen time
>> Whether we call this an attribute or a summary, it's basically still the same thing. And in my view at least, I think it would be a little more tricky to store the global property in a separate place to the local property, as they could more easily get out of sync.
>> 
>> Note that we do have plans to store a "summary" for LTO purposes in the bitcode file. That's essentially what Rafael has proposed with llvm.org/pr27551 <http://llvm.org/pr27551> -- we'd add a symbol table to the bitcode format that stores just the information needed to do symbol resolution against bitcode files. That symbol table would most likely include this property. But that work will take some time, and it doesn't solve the need for the information to be available at codegen time.
>> 
>> Thanks,
>> -- 
>> -- 
>> Peter
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits>_______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160531/06549162/attachment.html>