[llvm-dev] [RFC] Add IR level interprocedural outliner for code size.

[River Riddle via llvm-dev]

Hi Quentin,
 I appreciate the feedback. When I reference the cost of Target Hooks it's
mainly for maintainability and cost on a target author. We want to keep the
intrusion into target information minimized. The heuristics used for the
outliner are the same used by any other IR level pass seeking target
information, i.e TTI for the most part. I can see where you are coming from
with "having heuristics solely focused on code size do not seem realistic",
but I don't agree with that statement. I think there is a disconnect on
heuristics. The only user tunable parameters are the lower bound
parameters(to the cost model), the actual analysis(heuristic calculation)
is based upon TTI information.
 When you say "Would still be interesting to see how well this could
perform on some exact model (i.e., at the Machine level), IMO." I am
slightly confused as to what you mean. I do not intend to try and implement
this algorithm at the MIR level given that it exists in Machine Outliner.
There are several comparison benchmarks given in the "More detailed
performance data" of the original RFC. It includes comparisons to the
Machine Outliner when possible(I can't build clang on Linux with Machine
Outliner). I welcome any and all discussion on the placement of the
outliner in LLVM.
 Thanks,
River Riddle

On Mon, Jul 24, 2017 at 1:42 PM, Quentin Colombet <qcolombet at apple.com>
wrote:

> Hi River,
>
> On Jul 24, 2017, at 11:55 AM, River Riddle via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
> Hi Jessica,
>  The comparison to the inliner is an interesting one but we think it's
> important to note the difference in the use of heuristics. The inliner is
> juggling many different tasks at the same time, execution speed, code size,
> etc. which can cause the parameters to be very sensitive depending on the
> benchmark/platform/etc. The outliners heuristics are focused solely on the
> potential code size savings from outlining, and is thus only sensitive to
> the current platform. This only creates a problem when we are over
> estimating the potential cost of a set of instructions for a particular
> target. The cost model parameters are only minimums: instruction sequence
> length, estimated benefit, occurrence amount. The heuristics themselves are
> conservative and based upon all of the target information available at the
> IR level, the parameters are just setting a lower bound to weed out any
> outliers. You are correct in that being at the machine level, before or
> after RA, will give the most accurate heuristics but we feel there's an
> advantage to being at the IR level. At the IR level we can do so many more
> things that are either too difficult/complex for the machine level(e.g
> parameterization/outputs/etc). Not only can we do these things but they are
> available on all targets immediately, without the need for target hooks.
> The caution on the use of heuristics is understandable, but there comes a
> point when trade offs need to be made. We made the trade off for a loss in
> exact cost modeling to gain flexibility, coverage, and potential for
> further features. This trade off is the same made for quite a few IR level
> optimizations, including inlining. As for the worry about code size
> regressions, so far the results seem to support our hypothesis.
>
>
> Would still be interesting to see how well this could perform on some
> exact model (i.e., at the Machine level), IMO. Target hooks are cheap and
> choosing an implementation because it is simpler might not be the right
> long term solution.
> At the very least, to know what trade-off we are making, having prototypes
> with the different approaches sounds sensible.
> In particular, all the heuristics about cost for parameter passing
> (haven’t checked how you did it) sounds already complex enough and would
> require target hooks. Therefore, I am not seeing a clear win with an IR
> approach here.
>
> Finally, having heuristics solely focused on code size do not seem
> realistic. Indeed, I am guessing you have some thresholds to avoid
> outlining some piece of code too small that would end up adding a whole lot
> of indirections and I don’t like magic numbers in general :).
>
> To summarize, I wanted to point out that an IR approach is not as a clear
> win as you describe and would thus deserve more discussion.
>
> Cheers,
> -Quentin
>
>  Thanks,
> River Riddle
>
> On Mon, Jul 24, 2017 at 11:12 AM, Jessica Paquette <jpaquette at apple.com>
> wrote:
>
>>
>> Hi River,
>>
>> I’m working on the MachineOutliner pass at the MIR level. Working at the
>> IR level sounds interesting! It also seems like our algorithms are similar.
>> I was thinking of taking the suffix array route with the MachineOutliner in
>> the future.
>>
>> Anyway, I’d like to ask about this:
>>
>> On Jul 20, 2017, at 3:47 PM, River Riddle via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>> The downside to having this type of transformation be at the IR level is
>> it means there will be less accuracy in the cost model -  we can somewhat
>> accurately model the cost per instruction but we can’t get information on
>> how a window of instructions may lower. This can cause regressions
>> depending on the platform/codebase, therefore to help alleviate this there
>> are several tunable parameters for the cost model.
>>
>>
>> The inliner is threshold-based and it can be rather unpredictable how it
>> will impact the code size of a program. Do you have any idea as to how
>> heuristics/thresholds/parameters could be tuned to prevent this? In my
>> experience, making good code size decisions with these sorts of passes
>> requires a lot of knowledge about what instructions you’re dealing with
>> exactly. I’ve seen the inliner cause some pretty serious code size
>> regressions in projects due to small changes to the cost model/parameters
>> which cause improvements in other projects. I’m a little worried that an
>> IR-level outliner for code size would be doomed to a similar fate.
>>
>> Perhaps it would be interesting to run this sort of pass pre-register
>> allocation? This would help pull you away from having to use heuristics,
>> but give you some more opportunities for finding repeated instruction
>> sequences. I’ve thought of doing something like this in the future with the
>> MachineOutliner and seeing how it goes.
>>
>> - Jessica
>>
>>
> _______________________________________________
> 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/20170724/693b903d/attachment.html>