[llvm-dev] Aggregate load/stores

Tue Aug 18 22:31:06 PDT 2015

It is pretty clear people need this. Let's get this moving.

I'll try to sum up the point that have been made and I'll try to address
them carefully.

1/ There is no good solution for large aggregates.
That is true. However, I don't think this is a reason to not address
smaller aggregates, as they appear to be needed. Realistically, the
proportion of aggregates that are very large is small, and there is no
expectation that such a thing would map nicely to the hardware anyway (the
hardware won't have enough registers to load it all anyway). I do think
this is reasonable to expect a reasonable handling of relatively small
aggregates like fat pointers while accepting that larges ones will be
inefficient.

This limitation is not unique to the current discussion, as SROA suffer
from the same limitation.
It is possible to disable to transformation for aggregates that are too
large if this is too big of a concern. It should maybe also be done for
SROA.

2/ Slicing the aggregate break the semantic of atomic/volatile.
That is true. It means slicing the aggregate should not be done for
atomic/volatile. It doesn't mean this should not be done for regular ones
as it is reasonable to handle atomic/volatile differently. After all, they
have different semantic.

3/ Not slicing can create scalar that aren't supported by the target. This
is undesirable.
Indeed. But as always, the important question is compared to what ?

The hardware has no notion of aggregate, so an aggregate or a large scalar
ends up both requiring legalization. Doing the transformation is still
beneficial :
 - Some aggregates will generate valid scalars. For such aggregate, this is
100% win.
 - For aggregate that won't, the situation is still better as various
optimization passes will be able to handle the load in a sensible manner.
 - The transformation never make the situation worse than it is to begin
with.

On previous discussion, Hal Finkel seemed to think that the scalar solution
is preferable to the slicing one.

Is that a fair assessment of the situation ? Considering all of this, I
think the right path forward is :
 - Go for the scalar solution in the general case.
 - If that is a problem, the slicing approach can be used for non
atomic/volatile.
 - If necessary, disable the transformation for very large aggregates (and
consider doing so for SROA as well).

Do we have a plan ?

2015-08-18 18:36 GMT-07:00 Nicholas Chapman via llvm-dev <
llvm-dev at lists.llvm.org>:

> Oh,
> and another potential reason for handling aggregate loads and stores
> directly is that it expresses the semantics of the program more clearly,
> which I think should allow LLVM to optimise more aggresively.
> Here's a bug report showing a missed optimisation, which I think is due to
> the use of memcpy, which in turn is required to work around slow structure
> loads and stores:
> https://llvm.org/bugs/show_bug.cgi?id=23226
>
> Cheers,
>   Nick
>
> On 17/08/2015 22:02, mats petersson via llvm-dev wrote:
>
> I've definitely "run into this problem", and I would very much love to
> remove my kludges [that are incomplete, because I keep finding places where
> I need to modify the code-gen to "fix" the same problem - this is probably
> par for the course from a complete amateur compiler writer and someone that
> has only spent the last 14 months working (as a hobby) with LLVM].
>
> So whilst I can't contribute much on the "what is the right solution" and
> "how do we solve this", I would very much like to see something that allows
> the user of LLVM to use load/store withing things like "is my thing that
> I'm storing big, if so don't generate a load, use a memcpy instead". Not
> only does this make the usage of LLVM harder, it also causes slow
> compilation [perhaps this is a separte problem, but I have a simple program
> that copies a large struct a few times, and if I turn off my "use memcpy
> for large things", the compile time gets quite a lot longer - approx 1000x,
> and 48 seconds is a long time to compile 37 lines of relatively straight
> forward code - even the Pascal compiler on PDP-11/70 that I used at my
> school in 1980's was capable of doing more than 1 line per second, and it
> didn't run anywhere near 2.5GHz and had 20-30 users anytime I could use
> it...]
>
> ../lacsap -no-memcpy -tt longcompile.pas
> Time for Parse 0.657 ms
> Time for Analyse 0.018 ms
> Time for Compile 1.248 ms
> Time for CreateObject 48803.263 ms
> Time for CreateBinary 48847.631 ms
> Time for Compile 48854.064 ms
>
> compared with:
> ../lacsap -tt longcompile.pas
> Time for Parse 0.455 ms
> Time for Analyse 0.013 ms
> Time for Compile 1.138 ms
> Time for CreateObject 44.627 ms
> Time for CreateBinary 82.758 ms
> Time for Compile 95.797 ms
>
> wc longcompile.pas
>  37  84 410 longcompile.pas
>
> Source here:
> https://github.com/Leporacanthicus/lacsap/blob/master/test/longcompile.pas
>
>
> --
> Mats
>
> On 17 August 2015 at 21:18, deadal nix via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> OK, what about that plan :
>>
>> Slice the aggregate into a serie of valid loads/stores for non atomic
>> ones.
>> Use big scalar for atomic/volatile ones.
>> Try to generate memcpy or memmove when possible ?
>>
>>
>> 2015-08-17 12:16 GMT-07:00 deadal nix <deadalnix at gmail.com>:
>>
>>>
>>>
>>> 2015-08-17 11:26 GMT-07:00 Mehdi Amini <mehdi.amini at apple.com>:
>>>
>>>> Hi,
>>>>
>>>> On Aug 17, 2015, at 12:13 AM, deadal nix via llvm-dev <
>>>> llvm-dev at lists.llvm.org> wrote:
>>>>
>>>>
>>>>
>>>> 2015-08-16 23:21 GMT-07:00 David Majnemer <david.majnemer at gmail.com>:
>>>>
>>>>>
>>>>>
>>>>> Because a solution which doesn't generalize is not a very powerful
>>>>> solution.  What happens when somebody says that they want to use atomics +
>>>>> large aggregate loads and stores? Give them yet another, different answer?
>>>>> That would mean our earlier, less general answer, approach was either a
>>>>> bandaid (bad) or the new answer requires a parallel code path in their
>>>>> frontend (worse).
>>>>>
>>>>
>>>>
>>>> +1 with David’s approach: making thing incrementally better is fine *as
>>>> long as* the long term direction is identified. Small incremental changes
>>>> that makes things slightly better in the short term but drives us away of
>>>> the long term direction is not good.
>>>>
>>>> Don’t get me wrong, I’m not saying that the current patch is not good,
>>>> just that it does not seem clear to me that the long term direction has
>>>> been identified, which explain why some can be nervous about adding stuff
>>>> prematurely.
>>>> And I’m not for the status quo, while I can’t judge it definitively
>>>> myself, I even bugged David last month to look at this revision and try to
>>>> identify what is really the long term direction and how to make your (and
>>>> other) frontends’ life easier.
>>>>
>>>>
>>>>
>>> As long as there is something to be done. Concern has been raised for
>>> very large aggregate (64K, 1Mb) but there is no way a good codegen can come
>>> out of these anyway. I don't know of any machine that have 1Mb of register
>>> available to tank the load. Even I we had a good way to handle it in
>>> InstCombine, the backend would have no capability to generate something
>>> nice for it anyway. Most aggregates are small and there is no good excuse
>>> to not do anything to handle them because someone could generate gigantic
>>> ones that won't map nicely to the hardware anyway.
>>>
>>> By that logic, SROA should not exists as one could generate gigantic
>>> aggregate as well (in fact, SROA fail pretty badly on large aggregates).
>>>
>>> The second concern raised is for atomic/volatile, which needs to be
>>> handled by the optimizer differently anyway, so is mostly irrelevant here.
>>>
>>>
>>>>
>>>>>
>>>>
>>>>
>>>> clang has many developer behind it, some of them paid to work on it.
>>>> That s simply not the case for many others.
>>>>
>>>> But to answer your questions :
>>>>  - Per field load/store generate more loads/stores than necessary in
>>>> many cases. These can't be aggregated back because of padding.
>>>>  - memcpy only work memory to memory. It is certainly usable in some
>>>> cases, but certainly do not cover all uses.
>>>>
>>>> I'm willing to do the memcpy optimization in InstCombine (in fact,
>>>> things would not degenerate into so much bikescheding, that would already
>>>> be done).
>>>>
>>>>
>>>> Calling out “bikescheding” what other devs think is what keeps the
>>>> quality of the project high is unlikely to help your patch go through, it’s
>>>> probably quite the opposite actually.
>>>>
>>>>
>>>>
>>> I understand the desire to keep quality high. That's is not where the
>>> problem is. The problem lies into discussing actual proposal against
>>> hypothetical perfect ones that do not exists.
>>>
>>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org         http://llvm.cs.uiuc.edu
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>>
>
>
> _______________________________________________
> LLVM Developers mailing listllvm-dev at lists.llvm.org         http://llvm.cs.uiuc.eduhttp://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/20150818/3d54f3e0/attachment.html>