[llvm-dev] [Proposal][RFC] Epilog loop vectorization
Adam Nemet via llvm-dev
llvm-dev at lists.llvm.org
Tue Mar 14 10:11:46 PDT 2017
> On Mar 14, 2017, at 9:49 AM, Hal Finkel <hfinkel at anl.gov> wrote:
>
>
> On 03/14/2017 11:21 AM, Adam Nemet wrote:
>>
>>> On Mar 14, 2017, at 6:00 AM, Nema, Ashutosh <Ashutosh.Nema at amd.com <mailto:Ashutosh.Nema at amd.com>> wrote:
>>>
>>> Summarizing the discussion on the implementation approaches.
>>>
>>> Discussed about two approaches, first running ‘InnerLoopVectorizer’ again on the epilog loop immediately after vectorizing the original loop within the same vectorization pass, the second approach where re-running vectorization pass and limiting vectorization factor of epilog loop by metadata.
>>>
>>> <Approach-2>
>>> Challenges with re-running the vectorizer pass:
>>> 1) Reusing alias check result:
>>> When vectorizer pass runs again it finds the epilog loop as a new loop and it may generates alias check, this new alias check may overkill the gains of epilog vectorization.
>>> We should use the already computed alias check result instead of re computing again.
>>> 2) Rerun the vectorizer and hoist the new alias check:
>>> It’s not possible to hoist alias checks as its not fully redundant (not dominated by other checks), it’s not getting execute in all paths.
>>>
>>> <Mail Attachment.jpeg>
>>
>>
>> I don’t understand. Looks like you have the same alias checks for the epilog loop too. How is this CFG different from the re-vectorization of the scalar loop?
>
> You're looking at the wrong thing. This *is* the image from re-vectorization. The other image (linked below in step (3)) shows the other option.
Ah ok, the numbering confused me here.
>
>> Would be good to have both CFGs here and highlighting the difference.
>>
>> I thought that the whole point was that *if* you reached the epilog vector loop via the first vector loop, you want to bypass the alias checks before the epilog vector.
>
> Yes, but, that's not quite true now. You can also reach the epilogue loop if you fail the min-trip-count check, and so you don't know anything about the aliasing checks.
OK, so we want this loops to be handled specially. We effectively say that we only vectorize this loop if it does not require any alias checks or if the alias checks can be predicate-forwarded to this loop from existing checks.
This still seems like an orthogonal issue that may be interesting to solve independently. In other words this could be a nice feature in the vectorizer anyway: the loop is estimated to be low-trip count so feel free to predicate the new vector loop so that the alias check result could be reused from some other block. We obviously don’t have this capability today but it’s something that could be nice aside from the vectorizer.
Adam
>
> -Hal
>
>>
>> I still don’t understand why that’s not possible with some sophisticated predicate propagation independent from the vectorizer. I am not saying it’s already possible but it should be.
>>
>> Adam
>>
>>
>>>
>>> NOTE: We cannot prepone alias check as its expensive compared to other checks.
>>>
>>> <Approach-1>
>>> 1) Current patch depends on the existing functionality of LoopVectorizer, it uses ‘InnerLoopVectorizer’ again to vectorize the epilog loop, as it happens in the same vectorization pass we have flexibility to reuse already computed alias result check & limit vectorization factor for the epilog loop.
>>> 2) It does not generate the blocks for new block layout explicitly, rather it depends on ‘InnerLoopVectorizer::createEmptyLoop’ to generate new block layout. The new block layout get automatically generated by calling the ‘InnerLoopVectorizer:: vectorize’ again.
>>> 3) Block layout description with epilog loop vectorization is available at
>>> https://reviews.llvm.org/file/data/fxg5vx3capyj257rrn5j/PHID-FILE-x6thnbf6ub55ep5yhalu/LayoutDescription.png <https://reviews.llvm.org/file/data/fxg5vx3capyj257rrn5j/PHID-FILE-x6thnbf6ub55ep5yhalu/LayoutDescription.png>
>>>
>>> Approach-1 looks feasible, please comment if any objections.
>>>
>>> Regards,
>>> Ashutosh
>>>
>>>
>>> From: Nema, Ashutosh
>>> Sent: Wednesday, March 1, 2017 10:42 AM
>>> To: 'Daniel Berlin' <dberlin at dberlin.org <mailto:dberlin at dberlin.org>>
>>> Cc: anemet at apple.com <mailto:anemet at apple.com>; Hal Finkel <hfinkel at anl.gov <mailto:hfinkel at anl.gov>>; Zaks, Ayal <ayal.zaks at intel.com <mailto:ayal.zaks at intel.com>>; Renato Golin <renato.golin at linaro.org <mailto:renato.golin at linaro.org>>; mkuper at google.com <mailto:mkuper at google.com>; Mehdi Amini <mehdi.amini at apple.com <mailto:mehdi.amini at apple.com>>; llvm-dev <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>>
>>> Subject: RE: [llvm-dev] [Proposal][RFC] Epilog loop vectorization
>>>
>>> Sorry I misunderstood, gvn/newgvn/gvnhoist cannot help here as these checks are not dominated by all paths.
>>>
>>> Regards,
>>> Ashutosh
>>>
>>> From: Daniel Berlin [mailto:dberlin at dberlin.org <mailto:dberlin at dberlin.org>]
>>> Sent: Tuesday, February 28, 2017 6:58 PM
>>> To: Nema, Ashutosh <Ashutosh.Nema at amd.com <mailto:Ashutosh.Nema at amd.com>>
>>> Cc: anemet at apple.com <mailto:anemet at apple.com>; Hal Finkel <hfinkel at anl.gov <mailto:hfinkel at anl.gov>>; Zaks, Ayal <ayal.zaks at intel.com <mailto:ayal.zaks at intel.com>>; Renato Golin <renato.golin at linaro.org <mailto:renato.golin at linaro.org>>; mkuper at google.com <mailto:mkuper at google.com>; Mehdi Amini <mehdi.amini at apple.com <mailto:mehdi.amini at apple.com>>; llvm-dev <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>>
>>> Subject: Re: [llvm-dev] [Proposal][RFC] Epilog loop vectorization
>>>
>>> Hoisting or removing?
>>> Neither pass does hoisting, you'd need gvnhoist for that.
>>>
>>> Even then:
>>> Staring at your example, none of the checks are fully redundant (IE they are not dominated by other checks) or execute on all paths.
>>>
>>> Thus, hoisting them would be purely speculative code motion, which none of our passes do.
>>>
>>> If you would like these sets of checks to be removed, you would need to place them in a place that they execute unconditionally.
>>>
>>> Otherwise, this is not a standard code hoisting/removal transform.
>>>
>>> The only redundancy i can see here at all is the repeated getelementptr computation.
>>> If you move it to the preheader, it will be eliminated.
>>> Otherwise, none of the checks are redundant.
>>>
>>> What would you hope to happen in this case?
>>>
>>> On Tue, Feb 28, 2017 at 5:09 AM, Nema, Ashutosh <Ashutosh.Nema at amd.com <mailto:Ashutosh.Nema at amd.com>> wrote:
>>> I have tried running both gvn and newgvn but it did not helped in hoisting the alias checks:
>>>
>>> Please check, maybe I have missed something.
>>>
>>> <TestCase>
>>> void foo (char *A, char *B, char *C, int len) {
>>> int i = 0;
>>> for (i=0 ; i< len; i++)
>>> A[i] = B[i] + C[i];
>>> }
>>>
>>> <Command>
>>> $ opt –O3 –gvn test.ll –o test.opt.ll
>>> $ opt –O3 –newgvn test.ll –o test.opt.ll
>>>
>>> “test.ll” is attached, it got already vectorized by the approach running vectorizer twice by annotate the remainder loop with metadata to limit the vectorization factor for epilog vector loop.
>>>
>>> Regards,
>>> Ashutosh
>>>
>>> From: anemet at apple.com <mailto:anemet at apple.com> [mailto:anemet at apple.com <mailto:anemet at apple.com>]
>>> Sent: Tuesday, February 28, 2017 1:33 AM
>>> To: Hal Finkel <hfinkel at anl.gov <mailto:hfinkel at anl.gov>>
>>> Cc: Daniel Berlin <dberlin at dberlin.org <mailto:dberlin at dberlin.org>>; Nema, Ashutosh <Ashutosh.Nema at amd.com <mailto:Ashutosh.Nema at amd.com>>; Zaks, Ayal <ayal.zaks at intel.com <mailto:ayal.zaks at intel.com>>; Renato Golin <renato.golin at linaro.org <mailto:renato.golin at linaro.org>>; mkuper at google.com <mailto:mkuper at google.com>; Mehdi Amini <mehdi.amini at apple.com <mailto:mehdi.amini at apple.com>>; llvm-dev <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>>
>>> Subject: Re: [llvm-dev] [Proposal][RFC] Epilog loop vectorization
>>>
>>>
>>> On Feb 27, 2017, at 12:01 PM, Hal Finkel <hfinkel at anl.gov <mailto:hfinkel at anl.gov>> wrote:
>>>
>>>
>>> On 02/27/2017 01:47 PM, Daniel Berlin wrote:
>>>
>>>
>>> On Mon, Feb 27, 2017 at 11:29 AM, Adam Nemet <anemet at apple.com <mailto:anemet at apple.com>> wrote:
>>>
>>> On Feb 27, 2017, at 10:11 AM, Hal Finkel <hfinkel at anl.gov <mailto:hfinkel at anl.gov>> wrote:
>>>
>>>
>>> On 02/27/2017 11:47 AM, Adam Nemet wrote:
>>>
>>> On Feb 27, 2017, at 9:39 AM, Daniel Berlin <dberlin at dberlin.org <mailto:dberlin at dberlin.org>> wrote:
>>>
>>>
>>>
>>> On Mon, Feb 27, 2017 at 9:29 AM, Adam Nemet <anemet at apple.com <mailto:anemet at apple.com>> wrote:
>>>
>>> On Feb 27, 2017, at 7:27 AM, Hal Finkel <hfinkel at anl.gov <mailto:hfinkel at anl.gov>> wrote:
>>>
>>>
>>> On 02/27/2017 06:29 AM, Nema, Ashutosh wrote:
>>> Thanks for looking into this.
>>>
>>> 1) Issues with re running vectorizer:
>>> Vectorizer might generate redundant alias checks while vectorizing epilog loop.
>>> Redundant alias checks are expensive, we like to reuse the results of already computed alias checks.
>>> With metadata we can limit the width of epilog loop, but not sure about reusing alias check result.
>>> Any thoughts on rerunning vectorizer with reusing the alias check result ?
>>>
>>> One way of looking at this is: Reusing the alias-check result is really just a conditional propagation problem; if we don't already have an optimization that can combine these after the fact, then we should.
>>>
>>> +Danny
>>>
>>> Isn’t Extended SSA supposed to help with this?
>>>
>>> Yes, it will solve this with no issue already. GVN probably does already too.
>>>
>>> even if if you have
>>>
>>> if (a == b)
>>> if (a == c)
>>> if (a == d)
>>> if (a == e)
>>> if (a == g)
>>>
>>>
>>> and we can prove a ... g equivalent, newgvn will eliminate them all and set all the branches true.
>>>
>>> If you need a simpler clean up pass, we could run it on sub-graphs.
>>>
>>> Yes we probably don’t want to run a full GVN after the “loop-scheduling” passes.
>>>
>>> FWIW, we could, just without the memory-dependence analysis enabled (i.e. set the NoLoads constructor parameter to true). GVN is pretty fast in that mode.
>>>
>>> OK. Another data point is that I’ve seen cases in the past where the alias checks required for the loop passes could enable GVN to remove redundant loads/stores. Currently we can only pick these up with LTO when GVN is rerun.
>>>
>>> This is just GVN brokenness, newgvn should not have this problem.
>>> If it does, i'd love to see it.
>>>
>>> I thought that the problem is that we just don't run GVN after that point in the pipeline.
>>>
>>> Yeah, that is the problem but I think Danny misunderstood what I was trying to say.
>>>
>>> This was a datapoint to possibly rerun GVN with memory-awareness.
>>>
>>>
>>>
>>> -Hal
>>>
>>>
>>> (I'm working on the last few parts of turning it on by default, but it requires a new getModRefInfo interface to be able to get the last few testcases)
>>>
>>>
>>>
>>> --
>>> Hal Finkel
>>> Lead, Compiler Technology and Programming Languages
>>> Leadership Computing Facility
>>> Argonne National Laboratory
>>
>
> --
> Hal Finkel
> Lead, Compiler Technology and Programming Languages
> Leadership Computing Facility
> Argonne National Laboratory
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170314/d4f64789/attachment-0001.html>
More information about the llvm-dev
mailing list