[PATCH] Emit lifetime start/end for unnamed objects --- take 3

[David Blaikie]

On Tue, Sep 16, 2014 at 1:29 AM, Arnaud A. de Grandmaison <
arnaud.degrandmaison at arm.com> wrote:

> Hi David,
>
>
>
> In principal, giving more lifetime info can only improve stack slot
> sharing and reduce runtime stack usage, which is critical for us in the
> embedded world.
>
>
>
> I did not test that on a wide code base yet, but we had a customer
> reporting an issue where llvm/clang was producing code with a stack usage
> significantly worse than gcc. To make things worse, in their case their
> code was heavily recursive, to the point where using clang was simply not
> an option: they are forced to use gcc L
>
>
>
> This patch only adds lifetime markers to big enough (>32 bytes) objects,
> consistent with what is done for named temporaries. I do not know how this
> 32 bytes threshold has been choosen, but there is for sure a compile time /
> stack size gain trade-off to be made. My experiments have shown that for
> our customer case, the threshold should be lower: 16-bytes. But changing
> this threshold would require a separate thread on this list, as well as
> much more measurements.
>
>
>
> The improvements I have been able to get, by visual inspection of the
> generated assembly code, for a single call of the hot functions were:
>
>
>
>    | GCC | Clang | LT-32 | LT-16 |
>
> ===+=====+=======+=======+=======+
>
> F1 | 432 |   608 |   608 |   400 |
>
> F2 | 432 |   640 |   640 |   432 |
>
> F3 | 384 |   368 |   368 |   192 |
>
> F4 | 320 |   400 |   400 |   224 |
>
>
>
> Stack size is expressed in bytes.
>
> GCC version 4.8
>
> LT-32 is clang with this patch (default 32 bytes threshold for all
> temporaries).
>
> LT-16 is clang with this patch and a 16 bytes threshold for all
> temporaries.
>
>
>
> I believe bootstrapping clang could be a good testcase and will be needed
> when we will address the real problem in a separate discussion: what
> threshold should we use ?
>
>
>
> Very strangely to me coming from the embedded world, I have not found how
> to measure a program stack usage on linux, so if you have any idea, I am
> glad to hear about it.
>

I'm not sure what the nicest way to do it for the running program or
examining a binary, but I would /imagine/ that LLVM might have a
counter/statistic for stack usage. I believe LLVM has some way to record
statistics about optimizations, etc, for debugging the compiler. So if it
doesn't have a "stack size" stat counter, it could hopefully be added.

But you're right, for now - adding more should be generally better. I'm not
sure if there's a concern that adding too many (given that there's a
threshold, I assume someone tried it without a threshold and found that it
created too much metadata) intrinsics like this - so there might be some
need to show cost/benefit... (maybe looking at the commit archive to find
the original commits, those that added the threshold, etc)

(this is way out of my depth/area of interest - but just replying with
suggestions/ideas to both make the conversation more visible, give you some
ideas that might pre-empt ideas from other more knowledgeable reviewers,
etc)

- David


>
>
> Cheers,
>
> Arnaud
>
>
>
>
>
> *From:* David Blaikie [mailto:dblaikie at gmail.com]
> *Sent:* 16 September 2014 01:25
> *To:* Arnaud De Grandmaison
> *Cc:* llvm cfe
> *Subject:* Re: [PATCH] Emit lifetime start/end for unnamed objects ---
> take 3
>
>
>
> I'm hardly an expert on this stuff - but just curious: what sort of
> testing did you put this through? Bootstrap Clang? Were you able to gather
> any stats on reduced stack usage with this improvement to lifetime markers?
>
>
>
> On Mon, Sep 15, 2014 at 4:05 PM, Arnaud A. de Grandmaison <
> arnaud.degrandmaison at arm.com> wrote:
>
> Hi All,
>
>
>
> Please find attached a patch which teaches clang to emit lifetime.start /
> lifetime.end markers for unnamed temporary objects.
>
>
>
> This patch can greatly reduce the stack usage of some C++ code, where it
> is so easy to have short lived unnamed temporaries.
>
>
>
> As noted in the subject, this is my third attempt: my previous attempts
> failed to handle correctly the lifetime extended temporaries, and I have
> had a hard time to understand the CleanupScope. It all boiled down to the
> fact that the body of a function is not considered a full CleanupScope (for
> debug information reasons), so in the case of lifetime extended objects at
> the top level of the function body, with a trivial destructor  +
> lifetime.end marker, the lifetime markers were simply not considered,
> firing an assert in ~CodeGenFunction. All cases are now covered by
> testcases.
>
>
>
> I would appreciate if someone knowledgeable with the lifetime extended
> temporaries & cleanup scopes could give a look to this patch.
>
>
>
> Cheers,
>
> --
>
> Arnaud A. de Grandmaison
>
>
>
>
> _______________________________________________
> cfe-commits mailing list
> cfe-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20140916/8775081d/attachment.html>