llvm dev - Jul 2015 - [LLVMdev] AliasAnalysis update interface - a tale of sorrow and woe

----- Original Message -----
> From: "Hal Finkel" <hfinkel at anl.gov>
> To: "Chandler Carruth" <chandlerc at gmail.com>
> Cc: "LLVM Developers Mailing List" <llvmdev at
cs.uiuc.edu>, "Daniel Berlin" <dannyb at google.com>
> Sent: Thursday, July 2, 2015 3:14:38 PM
> Subject: Re: AliasAnalysis update interface - a tale of sorrow and woe
> 
> ----- Original Message -----
> > From: "Chandler Carruth" <chandlerc at gmail.com>
> > To: "LLVM Developers Mailing List" <llvmdev at
cs.uiuc.edu>, "Hal
> > Finkel" <hfinkel at anl.gov>, "Daniel Berlin"
> > <dannyb at google.com>
> > Sent: Wednesday, July 1, 2015 3:18:33 AM
> > Subject: AliasAnalysis update interface - a tale of sorrow and woe
> > 
> > 
> > Greetings folks,
> > 
> > 
> > As I'm working on refactoring the AA interfaces in LLVM to prepare
> > for the new pass manager, I keep hitting issues. Some of the
> > complexity that is hitting this stems from the update API, which
> > consists of three virtual functions:
> > 
> > 
> > deleteValue(Value *V)
> > copyValue(Value *From, Value *To)
> > addEscapingUse(Use &U)
> > 
> > 
> > These interfaces are *very* rarely called. Here are the only passes
> > that ever bother to use these:
> > - MemoryDependenceAnalysis.cpp
> > - BasicBlockUtils.cpp
> > - LoopSimplify.cpp
> > - MergedLoadStoreMotion.cpp
> > - GVN.cpp
> > 
> > 
> > That's it. This is not a complete set of code that can delete a
> > load
> > or store, nor a complete list of things which can fold something
> > into a Phi node or otherwise trigger an "escaping" use.
Maybe these
> > interfaces used to be much more widely used prior to the
> > introduction of AliasSetTracker? Now that utility is used instead.
> > Either way, basic things like CSE, store->load forwarding in
> > instcombine, etc., all seem like they *should* be updating AA, but
> > they aren't....
> > 
> > 
> > So, if these update APIs are really important, I'm pretty sure
that
> > LLVM is completely broken already...
> > 
> > 
> > But in fact, almost nothing overrides these APIs. The only pass in
> > tree that uses them at all is GlobalsModRef, which might explain
> > why
> > the world hasn't melted due to us not calling these update
routines
> > reliably.
> > 
> > 
> > So I looked into GlobalsModRef to understand why it is overriding
> > these. It doesn't override copyValue at all. That API point
appears
> > to be completely dead. So my first question is: can I remove
> > copyValue? Is there some out of tree user that desperately needs
> > this? If so, I'd like to understand why.
> > 
> > 
> > The APIs it does override are deleteValue to nix things in its
> > cache,
> > and addEscapingUse, which it implements by just deleting the used
> > value from its cache. This last one seems the most interesting.
> > Only
> > GVN calls it, but GVN does call it. So I commented out the code in
> > addEscapingUse so that it became a no-op. No test failed. =/ So I
> > added an assert to its addEscapingUse method. No test failed. So I
> > added a report_fatal_error to the addEscapingUse method and did an
> > LTO run over clang's bitcode which finally did reach this code
> > path.
> > 
> > 
> > addEscapingUse was added in 2011 by r122787 without any test case.
> > There is no mention of this fixing a bug. It looks like it may have
> > been intended to support something thata was never added to the
> > tree.
> > 
> > 
> > So I'd like to remove addEscapingUse since we used to not have it,
> > and we've never bothered to test it and I can't get anything
to
> > fail
> > without it, and GlobalsModRef is the only user and that is only
> > reached during LTO. Thoughts?
> > 
> > 
> > My final question is deleteValue. Again, only GlobalsModRef
> > implements this extension point. This was added in 2006 in r30684,
> > also without any changes to any test. It claims to implement a test
> > file, but that test isn't changed in the commit. Indeed,
commenting
> > out the body of deleteValue in GlobalsModRef causes no test to
> > fail.
> > =/ Fortunately, putting an assert here *does* trip in the
> > regression
> > test suite, so the code is reached, just not exercised.
> > 
> > 
> > Either way, the use case for the deleteValue at least makes perfect
> > sense. But there is (IMO) a much better way to accomplish the same
> > task: use a ValueHandle to trigger the update on deletion. This
> > will
> > prevent widespread failure to use the deleteValue API to update
> > alias analysis. So I would like to replace deleteValue with a
> > ValueHandle approach. But I'm more than a little nervous
> > implementing this, *as no test actually uses the behavior*! At
> > least
> > the code is reached...
> > 
> > 
> > I think generally, the update API in the AliasAnalysis interface
> > isn't working well. It is poorly and incompletely used. It
isn't
> > even clear that it is necessary in the face of tools like value
> > handles. Thoughts about just removing the entire thing and falling
> > back to value handles everywhere?
> 
> As a side point, I don't think we currently have a ValueHandle type
> that can detect when a new use of a value is added. Do we? This
> would certainly be useful for some kind of AA result caching, etc.
I should add that I'm somewhat afraid, however, of the compile-time
implications of adding the necessary hook.

 -Hal
> 
>  -Hal
> 
> > 
> > 
> > -Chandler
> 
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
> 
-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory

> On Jul 2, 2015, at 1:17 PM, Hal Finkel <hfinkel at anl.gov> wrote:
> 
> I should add that I'm somewhat afraid, however, of the compile-time
implications of adding the necessary hook.
I was just going to say this myself.

In fact, even if you don’t add a hook for tracking new values, just using value
handles in AA could prove to be expensive.

A typical optimized piece of code has more load instructions than binary
instructions.  If you only cached loads with value handles, you’re already
creating a huge number of handles, and these are currently entries in a map
(pImpl->ValueHandles[V] if you’re curious).

TL;DR: If we are going to create this many value handles, we should strongly
consider finding a way to represent them more cheaply.

Cheers,
Pete

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150702/14ecb447/attachment.html>

The vast majority of stateful AA probably doesn't/won't care about the
vast majority of updates.

For example, unless globalsmodref is going to recompute the transitive
closure, it can't actually give better than "i don't know"
answers in
most cases, which is equivalent to telling it "you should delete this
pointer and pretend you know nothing about it".  Humorously, you can
see that's what it did with addEscapingUse. It deleted anything it
knew about the pointer.

Further, once that delete has happened once, there is no point in ever
tracking anything about that value ever again for globalsmodref.  It
will never regain that info.

Even for those that do care, there are other issues:
The AA may never be queried again.  Or the IR may change 6 or 7 times
before it gets queried again.  The only thing any AA needs to know in
general is "what is the net result of those changes" (IE added uses,
etc). For example, if you were to add and then remove a use before
querying AA, no AA will care. Telling it an add and then remove is not
only pointless and expensive, it is probably worse than telling it
"nothing happened" (which is the net effect).

In any case, i think we need to sit down and think "which of these
AA's do we want to cache/make stateful, and how".
Because the original use case for building a stateful API, from what i
can tell, is globalsmodref.  This is not a great use case, since
incremental updating may require a complete callgraph walk.  That
seems like a non-starter :-)

Nowadays, i expect the stateful use cases are more like "caching
CFL-AA or caching expensive parts of BasicAA" (both of which are
O(pointer chain) to update incrementally).
Even andersen's can be incrementally computed quicker in practice (For
additional uses, delete variable from sets, add new constraints,
fixpoint.  In most cases, you can prove the new constraint doesn't
change the results, so you do nothing.  Or it takes one iteration of
propagation.  Worst case is N^3 like the callgraph walk, but in
practice, it's probably closer to O(pointer chain))





On Thu, Jul 2, 2015 at 2:05 PM, Pete Cooper <peter_cooper at apple.com>
wrote:
>
> On Jul 2, 2015, at 1:17 PM, Hal Finkel <hfinkel at anl.gov> wrote:
>
> I should add that I'm somewhat afraid, however, of the compile-time
> implications of adding the necessary hook.
>
> I was just going to say this myself.
>
> In fact, even if you don’t add a hook for tracking new values, just using
> value handles in AA could prove to be expensive.
>
> A typical optimized piece of code has more load instructions than binary
> instructions.  If you only cached loads with value handles, you’re already
> creating a huge number of handles, and these are currently entries in a map
> (pImpl->ValueHandles[V] if you’re curious).
>
> TL;DR: If we are going to create this many value handles, we should
strongly
> consider finding a way to represent them more cheaply.
>
> Cheers,
> Pete
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>