[llvm-dev] LLD status update and performance chart

[Rui Ueyama via llvm-dev]

On Fri, Dec 16, 2016 at 12:31 PM, Pete Cooper <peter_cooper at apple.com>
wrote:

>
> On Dec 16, 2016, at 11:46 AM, Rui Ueyama <ruiu at google.com> wrote:
>
> On Fri, Dec 16, 2016 at 11:18 AM, Pete Cooper <peter_cooper at apple.com>
> wrote:
>
>> Hi Rui
>>
>> I agree separating the components out in to libraries only makes sense
>> when there is a clear reason to do so.  However, just this year there was a
>> very involved discussion about what it means to be a library.
>> Specifically, I don't think your current 'main-as-library' argument is
>> valid while you call exit or (if you) rely on mutable global state.  Having
>> a single entry point via a main function is fine, but that function cannot
>> then kill the process which its linked in to.
>>
>
> Our main function returns as long as input object files are not corrupted.
> If you are doing in-memory linking, I think it is unlikely that the object
> files in memory are corrupted (especially when you just created them using
> LLVM), so I think this satisfies most users needs in practice. Do you have
> a concern about that?
>
> Ultimately my concern is that there is *any* code path calling exit.  I
> would say that this prevents the lld library from being used in-process.
> But others opinions may differ, and I honestly don't have a use case in
> mind, just that I don't think library code should ever call exit.
>

There is a duality of LLD: lld-as-a-command and lld-as-a-library. This
duality is not necessarily a bad thing. Given that we have a verifier, any
path that leads to check for an impossible error condition and call exit()
should be thought as an assert() when it is used as a library since they
should never happen or there is a bug in code (and that's what assert
actually does). We already have lots of asserts in our libraries, and
that's I think essentially the same.


> For the situation that you need to handle foreign object files in the same
> process (I'd recommend you to sandbox a process in that case though), we
> can write a verifier to check for file correctness rigorously so that we
> can guarantee that object files are as trustworthy as freshly-created
> object files. I think this feature is a reasonable addition to the linker.
>
> That sounds great.  Having written some parts of the MachO lld linker and
> seen Kevin's work on llvm-objdump, I can appreciate that is not easy.  For
> example, I wrote the logic to process EH FDE's which may need to error out
> if invalid.  You don't necessarily want to validate them all up front as it
> may be too slow, so I can understand that this isn't necessarily trivial to
> handle in a performant way.
>

That's I don't know yet. My gut is that doing error checking beforehand
makes code easy to read and maintain, just like semantic analysis doesn't
have to handle syntactic errors. But I don't know the answer, so I cannot
exclude neither possibilities. We have to experiment that and compare.


>
> As to the mutable shared state, my current (unproved) idea is to make them
> thread local variables. Since no one yet has come up to say "hey, we are
> actually trying to run multiple instances of the linker in the same process
> simultaneously but LLD doesn't allow that", that's not implemented yet, but
> technically I think it's doable, and that's needless to say a reasonable
> feature request.
>
> LLVM uses the LLVMContext for this (and begs users to look the other way
> with regards to cl::opt's).  I don't know if there's been a discussion in
> LLVM about whether TLV's would be better there too, but seems like a
> reasonable discussion to have.  Certainly I don't think anyone should say
> you can't use them without good reason.
>

That's also another thing no one knows the answer. As far as I can say,
global states in the LLD/ELF makes things easy to maintain, and looks like
a majority of people working on it are in favor of it. Of course people who
have different taste may not like it that much, I understand that, and I
don't say that that's the best way, but it's there and it works fairly
satisfactory. The most important thing for external users is the API, no?
We can discuss what is the best way to have a linker-global state
internally, but as long as we provide a sane API, everything else should
fall in the internal design stuff category.


>
> As I repeatedly said in the thread that speed is not the only goal for us.
> Honestly, it's going to be the best selling point of LLD, because most
> people do not use that many linker features but just use it to create
> executables (and sometimes wait for a long period of time). I reported
> about the performance in this thread because I thought people would be
> happy to hear the speed improvement we've made this year. Also, because I
> was happy about that, I probably emphasized that too much. But that's not
> our single goal.
>
> I meant to commend you for both sending out a summary email, and the
> results.  Having this fast a linker on ELF/COFF is going to be a huge win
> for developers.  And I personally really like status updates for major
> projects/features as it can be hard to follow along with all the email
> traffic.  So thank you for doing that.
>
> My only concern with performance is that I felt like you would be against
> changes to the code which make it slower but add functionality.  Error
> handling is such a use case.  LLVM and clang continue to get bigger each
> year and sometimes that means a little slower too.  The linker may be
> faster next year than it is now, or it may be slower but have a feature
> which makes that a worthwhile tradeoff.  I don't want to slow down any of
> the code for any reason, but its natural that sometimes it'll happen with
> good reason.
>

I don't know if you believe me by repeating the thing I said many times in
this thread, but I did not sacrificing functionality for speed.

If you take a look at the performance chart that I sent in this thread,
you'll notice the pattern that the linker gradually became slower and then
suddenly became faster. As we add more safety measures, error checks and
features, the linker get slower and slower. Each one is small but they
accumulates. And then we sometimes ran a profiler to nail down a
bottleneck, came up with a good optimization, and implement it. That's what
you see as steep speedups in the chart. We do not optimize by removing
features. We just did better.

So, I don't know what I can do for you to believe me, but I have never said
that the performance is the only goal, and you can find it by my actual
behavior. I believe I've been trying to always be helpful. Please ask
LLD/ELF developers about that. If you find me doing the opposite in the
code review or discussion, please point it out, so that I can correct that.


> Thanks,
> Pete
>
>
> If you want  context then the relevant piece of the thread is
>> http://lists.llvm.org/pipermail/llvm-dev/2016-January/093760.html.
>>
>> Arseny summarized things very well there, so i'll just quote him at the
>> end here.  I understand that you and others want to first write a fast
>> linker tool and i don't think anyone has any problem with that, but there
>> is also a clear desire from folks to have it be usable as a library and I
>> would hope any patches to do so are accepted, even if they make the code
>> more complex, or slower.
>>
>> >>>* On Thu, Jan 7, 2016 at 7:03 AM, Arseny Kapoulkine via llvm-dev <
>> *>>>* llvm-dev at lists.llvm.org <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>> wrote:
>> *>>>>>>>* In the process of migrating from old lld ELF linker to new (previously
>> *>>>>* ELF2) I noticed the interface lost several important features (ordered by
>> *>>>>* importance for my use case):
>> *>>>>>>>>* 1. Detecting errors in the first place. New linker seems to call
>> *>>>>* exit(1) for any error.
>> *>>>>>>>>* 2. Reporting messages to non-stderr outputs. Previously all link
>> *>>>>* functions had a raw_ostream argument so it was possible to delay the error
>> *>>>>* output, aggregate it for multiple linked files, output via a different
>> *>>>>* format, etc.
>> *>>>>>>>>* 3. Linking multiple outputs in parallel (useful for test drivers) in a
>> *>>>>* single process. Not really an interface issue but there are at least two
>> *>>>>* global pointers (Config & Driver) that refer to stack variables and are
>> *>>>>* used in various places in the code.
>> *>>>>>>>>* All of this seems to indicate a departure from the linker being useable
>> *>>>>* as a library. To maintain the previous behavior you'd have to use a linker
>> *>>>>* binary & popen.*
>>
>> Pete
>>
>> On Dec 16, 2016, at 10:15 AM, Rui Ueyama via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>> I talked several people and found that this is more like a communication
>> issue rather than a technical/philosophical issue. I believe communication
>> problems won't solve themselves. As a person who is on the owners file of
>> LLD, I think I need to say something about that issue. Also, I guess people
>> who were just watching this thread wondered why my happy pre-holiday status
>> report suddenly turned into a heated discussion, and they are probably
>> still wondering what's wrong with LLD. I want to address that, too.
>>
>> So, as a project, there is no anti-library policy in LLD. I think this is
>> the misunderstanding one side had. We already provide main-as-a-library
>> feature so that you can embed the linker to your program. We as a project
>> welcome other ideas to export linker features at a well-defined boundary.
>> For example, I think abstracting the file system access so that you can
>> hook file operations could be a well-defined, useful API for those who want
>> to do in-memory linking (I expressed that opinion already in this thread).
>> Just like LLVM, we won't guarantee API compatibility between releases, and
>> we are unlikely to be able to expose deep internals of the linker, but as
>> long as you think you found a reasonable coarse API boundary, there should
>> be nothing preventing you from bringing that to the table.
>>
>> On the other hand, as far as I talked, no one who is on the "library"
>> side requested LLD expose deep internals. This is the misunderstanding the
>> other side had. If we as a project said that LLD should not support any
>> library interface at all, they would be upset and speak out loudly, but
>> again, that's not a project policy.
>>
>> So, correct me if I'm wrong, but I don't see no serious conflicts here.
>> The conflict I saw in the thread is I believe superficial, and I strongly
>> believe that it could have been addressed calmly and nicely if we have used
>> more words to explain thoughts instead of small number of strong words.
>>
>> Hope this helps.
>>
>> Rui
>>
>> On Fri, Dec 16, 2016 at 1:40 AM, George Rimar via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>>> >I am on PTO, so slow to respond.
>>> >
>>> >Some items that are left:
>>> >
>>> >* Debug fission
>>> >* Single file debug fission
>>> >* Range extension thunks
>>> >* All of freebsd links and works
>>> >* Very good performance when all that is in
>>>
>>> Looks we have initial version of debug fusion implemented.
>>> r289790, r289810 commits from yesterday did the rest of main job I
>>> believe.
>>> I do not know what is "Single file debug fission" ? (quick googling
>>> gives nothing and I never heard about that before I think)
>>>
>>> George.
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161216/3b7a12da/attachment-0001.html>