[llvm-dev] Possible soundness issue with available_externally (split from "RFC: Add guard intrinsics")

Xinliang David Li via llvm-dev llvm-dev at lists.llvm.org
Fri Feb 26 20:48:13 PST 2016

On Fri, Feb 26, 2016 at 8:43 PM, Chandler Carruth <chandlerc at google.com>

> On Fri, Feb 26, 2016 at 8:40 PM Xinliang David Li <xinliangli at gmail.com>
> wrote:
>> On Fri, Feb 26, 2016 at 8:30 PM, Chandler Carruth <chandlerc at google.com>
>> wrote:
>>> On Fri, Feb 26, 2016 at 8:15 PM Hal Finkel <hfinkel at anl.gov> wrote:
>>>> Can you explain your concerns about reliability? My idea is that we
>>>> have a kind of delayed-name comdat section attribute that says that the
>>>> name of the comdat section into which the function will be placed will be
>>>> based on the hash of the contents (which should probably be the pre-isel
>>>> IR-level contexts plus perhaps the other state that might affect backend
>>>> optimization). Maybe just using the MI-level function contents will work
>>>> too. It seems like this should be "perfect". I don't see the circumstances
>>>> why function definitions optimized under identical optimization settings
>>>> should differ in this sense.
>>> Ok, let's consider Sanjoy's latest post.
>>> I think we could end up with this routine in N different modules with
>>> slightly different amounts of information in each. Each of these in turn
>>> could cause us to deduce a slightly different set of attributes, and thus
>>> compute a different hash.
>>> Now, certainly, there are a small number of variants here. Two, maybe
>>> three. But this could in the worst case cause us to have two copies of a
>>> *huge* number of functions.
>>> My concern here is not that this is a *theoretically* hard problem to
>>> solve. I think there are good techniques to handle this. My concern is that
>>> it will be a substantial amount of work and require a reasonable amount of
>>> testing and experimentation.
>> Such data won't be too hard to find out. It can be conservatively
>> estimated by looking at the final binary code for each COMDAT to see how
>> many 'clones's are needed due to different code gen -- the result is the
>> upper bound as different code gen does not mean different deduced function
>> attributes.
>> With PGO, the decision process will also become much simpler.
> Sure. And if you want to work on privatization in LLVM's IPO, that would
> be awesome.
> But I think it is completely unreasonable to gate checking in code to test
> for the invalid transformation until this new optimization technique is
> available. We should be fixing the bug now, and then figuring out how to do
> privatization.

Isn't privatization one way to get correctness? Other techniques such as
hash based de-dup or use PGO etc for tuning are orthognal methods to reduce
the overall cost of the former.


> -Chandler
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160226/fd96d0dc/attachment.html>

More information about the llvm-dev mailing list