[LLVMdev] GlobalsModRef (and thus LTO) is completely broken

[Chandler Carruth]

I've fixed the obvious bugs I spotted in r242281. These should be pure
correctness improvements.

I've sent the two patches I'm imagining to address the core issue here:
http://reviews.llvm.org/D11213
http://reviews.llvm.org/D11214

Currently, I have the unsafe alias results disabled by default, but with a
flag that can re-enable them if needed. I don't feel really strongly about
which way the default is set -- but that may be because I don't have lots
of users relying on LTO. I'll let others indicate which way they would be
most comfortable.

Some IRC conversations indicated that early benchmark results with GMR
completely disabled weren't showing really significant swings, so maybe
this relatively small reduction in power of GMR won't be too problematic
for folks. Either way, I'm open to the different approaches. It's D11214
that I care a lot about. =]


Thanks for all the thoughts here!
-Chandler

On Tue, Jul 14, 2015 at 11:25 PM Chandler Carruth <chandlerc at gmail.com>
wrote:

> Replying here, but several of the questions raised boil down to "couldn't
> you make the usage of GetUnderlyingObject conservatively correct?". I'll
> try and address that.
>
> I think this *is* the right approach, but I think it is very hard to do
> without effectively disabling this part of GlobalsModRef. That is, the easy
> ways are likely to fire very frequently IMO.
>
> The core idea is to detect a "no information" state coming out of
> GetUnderlyingObject (likely be providing a custom version just for
> GlobalsModRef and tailored to its use cases). This is particularly
> effective at avoiding the problems with the recursion limit. But let's look
> at what cases we *wouldn't* return that. Here are the cases I see when I
> thought about this last night with Hal, roughly in descending likelihood I
> would guess:
>
> 1) We detect some global or an alloca. In that case, even BasicAA would be
> sufficient to provide no-alias. GMR shouldn't be relevant.
>
> 2) We detect a phi, select, or inttoptr, and stop there.
>
> 3) We detect a load and stop there.
>
> 4) We detect a return from a function.
>
> 5) We detect an argument to the function.
>
> I strongly suspect the vast majority of queries hit #1. That's why BasicAA
> is *so* effective. Both #4 and #5 I think are actually reasonable places
> for GMR to potentially say "no-alias" and provide useful definitive
> information. But I also suspect these are the least common.
>
> So let's look at #2 and #3 because I think they're interesting. For these,
> I think it is extremely hard to return "no-alias". It seems extremely easy
> for a reasonable and innocuous change to the IR to introduce a phi or a
> select into one side of the GetUnderlyingObject but not the other. If that
> ever happens, we can't return "no-alias" for #2, or we need to add really
> expensive updates. It also seems reasonable (if not as likely) to want
> adding a store and load to the IR to not trigger a miscompile. If it is
> possible for a valid optimization pass to do reg2mem on an SSA value, then
> that could happen to only one side of the paired GetUnderlyingObject and
>  break GMR with #3. If that seems like an unreasonable thing to do,
> consider loop re-rolling or other transformations which may need to take
> things in SSA form at move them out of SSA form. Even if we then try
> immediately to put it back *into* SSA form, before we do that we create a
> point where GMR cannot correctly return no-alias.
>
> So ultimately, I don't think we want to rely on GMR returning "no-alias"
> for either #2 or #3 because of the challenge of actually updating it in all
> of the circumstances that could break them. That means that *only* #4 and
> #5 are going to return "no-alias" usefully. And even then, function
> inlining and function outlining both break #4 and #5, so you have to
> preclude those transforms while GMR is active. And I have serious doubts
> about these providing enough value to be worth the cost.
>
>
> I think the better way to approach this is the other way around. Rather
> than doing a backwards analysis to see if one location reaches and global
> and the other location doesn't reach a global, I think it would be much
> more effective to re-cast this as a forward analysis that determines all
> the memory locations in a function that come from outside the function, and
> use that to drive the no-alias responses.
>
>
> On Tue, Jul 14, 2015 at 12:12 PM Gerolf Hoflehner <ghoflehner at apple.com>
> wrote:
>
>> I wouldn’t be willing to give up performance for hypothetical issues.
>> Please protect all your changes with options. For some of your concerns it
>> is probably hard to provide a test case that shows an/the actual issue.
>>
>
> I certainly agree that it will be very hard to provide a test case and
> extremely rare to see this in the wild for most of these issues. As long as
> I can remove the problematic update API we currently have (which as its an
> API change can't really be put behind flags), I'm happy to have flags
> control whether or not GMR uses the unsound / stale information to try to
> answer alias queries. Do you have any opinion about what the default value
> of the flags should be?
>
> I'll go ahead and prepare the patches, as it seems like we're all ending
> up in the same position, and just wondering about the precise tradeoffs we
> want to settle on.
>
>
>> _______________________________________________
>> LLVM Developers mailing list
>>
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150715/91f7cc76/attachment.html>