[llvm-dev] RFC: Strong GC References in LLVM

Daniel Berlin via llvm-dev llvm-dev at lists.llvm.org
Thu Jul 21 17:14:21 PDT 2016


On Thu, Jul 21, 2016 at 4:25 PM, Philip Reames <listmail at philipreames.com>
wrote:

> It sounds like this new analysis could basically subsume the existing
> OrderedBasicBlock.
>
> Thinking about update semantics, what types of updates need to be recorded
> to make this efficient?
> - Adding maythrow instructions is obvious required, figuring out how to
> update clearly requires OBB info.
>

Yup.


> - Adding arithmetic instructions - only requires OBB update
>

This is possible to do by skip numbering and using uint64.

If a given pass performs only insertafter calls, you can show that you can
add tons of instructions with O(1) updates by skip-numbering
Same with only insertbefore.
You can use the same numbering scheme for both, and just add/subtract 1 in
the right places until you run out of space.  With int64 as the numbers,
it's "a lot of places" :)

If you allow interspersing of insertbefore/insertafter, the numbering
scheme is more complex, and you can only insert max of log(n) instructions
between two instructions before you must renumber.

(this is just using previous inst number+next inst number/2 as the new
number)


Whether any of this is worth it,  ¯\_(ツ)_/¯

I'll probably start without it and we can do it if it's too slow.



> I'm thinking that some form of lazy update might be the best path forward
> here.  Have the results re-calculated via a linear scan of the block only
> once a) a change has happened and been marked and b) the request
> information is required.  This would allow batch updates for instance.  We
> could potentially try to incremental update the motion barriers, but that
> seems like potentially too much work.
>
Yup.




>
> Also, I'm now thinking that first and last isn't enough information.  Once
> I ask the speculation question, I think need to ask what the *next*
> ordering barrier is.  I thinking code of the following variety:
> for block I want to skip:
>   if not has ordering barrier:
>     continue
>   if can speculate over block
>     continue
>   barrier = last barrier in block
>   while can speculate over barrier:
>     barrier = previous barrier
>   return next(barrier) as insert pt
> return first instruction in function entry
>


So, so far nothing tries to do that.

It's easy enough to do though, we can just maintain unordered maps of
 inst, number, and ordered maps of number, inst (assuming we allow for
incremental update) per block.

Or at least, this is the "simple" way. You can do it in a better time query
time bound by unordered maps of inst, {number, next/last inst} (IE you give
us the current thing we gave you, we tell you the next or last thing from
it).
This is not sane to incrementally update though.




>
>
> On 07/21/2016 12:17 PM, Daniel Berlin wrote:
>
> Okay, so it sounds like it might actually be better to be even more low
> level, call it "ExtendedBBInfo" or something, and rename what it provides
> to be more clearly structural:
>
> A. Inst * FirstIsGuaranteedToTransferExecutionToSuccessor(BB) (naming
> bikeshed open on this one :P)
> B. Inst * LastIsGuaranteedToTransferExecutionToSuccessor(BB)
> C. Inst *FirstMayThrow(BB)
> D. Inst *LastMayThrow(BB)
>
>
> Most things want to know if a given inst is before or after these.
>
> Since we have to touch the entire set of instructions for a bb anyway, we
> could also provide dominates (like orderedbasicblock) to give you the
> answer to that question for free (otherwise everything has to rewalk the
> entire inst list again)
>
> Rather than make it part of the API for this class, this would basically
> be making OrderedBasicBlock an interface, and this class happens to be able
> to provide a pointer to something satisfying that interface as well.
>
> (IE getOrderedBasicBlock() on EBBInfo will return you something that  has
> locallyDominates())
>
>
>
>
> On Thu, Jul 21, 2016 at 11:59 AM, Philip Reames <listmail at philipreames.com
> > wrote:
>
>>
>>
>> On 07/21/2016 10:30 AM, Daniel Berlin wrote:
>>
>>
>>
>> On Thu, Jul 21, 2016 at 9:39 AM, Daniel Berlin < <dberlin at dberlin.org>
>> dberlin at dberlin.org> wrote:
>>
>>>
>>>
>>> On Thu, Jul 21, 2016 at 9:26 AM, Andrew Trick < <atrick at apple.com>
>>> atrick at apple.com> wrote:
>>>
>>>>
>>>> On Jul 21, 2016, at 7:45 AM, Philip Reames <
>>>> <listmail at philipreames.com>listmail at philipreames.com> wrote:
>>>>
>>>> Joining in very late, but the tangent here has been interesting (if
>>>> rather OT for the original thread).
>>>>
>>>> I agree with Danny that we might want to take a close look at how we
>>>> model things like maythrow calls, no return, and other implicit control
>>>> flow.  I'm not convinced that moving to a pure explicit model is the right
>>>> idea because we get a lot of gain in practice from being able to reason
>>>> about what are essentially a limited form of extended basic blocks.  I
>>>> would welcome a design discussion about this, preferably in person, but
>>>> also don't want to block any current (or future honestly) work on this
>>>> until we have a reasonable firm plan of action.
>>>>
>>>> One idea would be to explicitly acknowledge that our "basic blocks" are
>>>> actually "extended basic blocks" with internal exits due to exception
>>>> propagation, noreturn, and (recently) guards.  I could see a couple of
>>>> implementation strategies here:
>>>> 1) Extend BasicBlock to explicitly track potential early exiting
>>>> instructions.  This would probably have to be conservative so that things
>>>> like nothrow inference aren't required to update all callers in one go.
>>>> 2) Conservative assume that BasicBlock has an unknown number of early
>>>> exiting instructions and write an analysis pass which answers questions
>>>> about where those early exits are.  Any pass which does code motion would
>>>> require this analysis.  (This is essentially the principled version of our
>>>> current hacks.)
>>>>
>>>>
>>>> This analysis can be lazy/incremental. Most passes only do “safe”
>>>> speculation and code sinking without side effects.
>>>>
>>>
>>> While I agree it can be lazy, and should be an analysis, i'm, again,
>>> really not sure which passes you are thinking about here that do code
>>> sinking/speculation that won't need it.
>>>
>>> Here's the list definitely needing it right now:
>>> GVN
>>> GVNHoist
>>> LICM
>>> LoadCombine
>>> LoopReroll
>>> LoopUnswitch
>>> LoopVersioningLICM
>>> MemCpyOptimizer
>>> MergedLoadStoreMotion
>>> Sink
>>>
>>> The list is almost certainly larger than this, this was a pretty trivial
>>> grep and examination.
>>> (and doesn't take into account bugs, etc)
>>>
>>>
>>
>> (Note, this list is stuff in the Scalar directory only. Everything in
>> Vectorize would n eed it, etc)
>>
>> And in case folks want more evidence that reasonable llvm developers need
>> something that just gives easy and right answers, see
>> <https://reviews.llvm.org/D22547>https://reviews.llvm.org/D22547 from
>> just yesterday :)
>>
>> To move this along, i will go build the analysis (I already have it
>> mostly done, actually).  If someone updates our docs to make this stuff
>> clear and obvious, that would be wonderful :)
>>
>> The analysis currently computes, internally, for a given BB:
>>
>> EarliestLoadHoistBarrier (used to see if you can move something out of a
>> block)
>> LatestLoadHoistBarrier (used to find the latest safe insertion point in a
>> block, if any)
>> EarliestStoreSinkBarrier (insertion)
>> LatestStoreSinkBarrier (movement)
>>
>>
>> (stores are maythrow dependent, loads are
>> isGuaranteedToTransferExecutionToSuccessor dependent)
>>
>> I'm still coming up with the external API, the users all want to know
>> either:
>>
>> 1. Can i move a load up out of this block to a direct predecessor
>> 2. Can i move a store down out of this block to a direct successor
>>
>> I would argue that we should have this analysis *only* report the EBB
>> information.  Doing this in the form of the information you've mentioned is
>> fine, but I would not want to see this become our deref analysis for
>> instance.  I think that needs to be a separate analysis which is well
>> layered on top of this one.  (i.e. once I encounter a hoist barrier, I need
>> to then ask if a load is safe to speculate beyond that point as a separate
>> question.)
>>
>>
>> GVN's current load PRE is complex to get "right" from it's current
>> standpoint, the APi that will provide the easiest way to fix it will be:
>>
>> 3. What is the latest insertion point in a block for a load, if it is
>> safe (IE the block does not end in an instruction you cannot move the load
>> before).
>>
>> Nothing is doing global store sinking right now that i see, so nothing
>> needs the analogous store api.
>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160721/2373a164/attachment.html>


More information about the llvm-dev mailing list