<div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote">On Sun, Jan 25, 2015 at 7:57 AM, Hal Finkel <span dir="ltr"><<a href="mailto:hfinkel@anl.gov" target="_blank">hfinkel@anl.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="HOEnZb"><div class="h5">----- Original Message -----<br>
> From: "Chandler Carruth" <<a href="mailto:chandlerc@google.com">chandlerc@google.com</a>><br>
> To: "LLVM Developers Mailing List" <<a href="mailto:llvmdev@cs.uiuc.edu">llvmdev@cs.uiuc.edu</a>>, "Daniel Berlin" <<a href="mailto:dannyb@google.com">dannyb@google.com</a>>, "Hal Finkel"<br>
> <<a href="mailto:hfinkel@anl.gov">hfinkel@anl.gov</a>>, "George Burgess IV" <<a href="mailto:george.burgess.iv@gmail.com">george.burgess.iv@gmail.com</a>>, <a href="mailto:apazos@codeaurora.org">apazos@codeaurora.org</a>, "Nick Lewycky"<br>
> <<a href="mailto:nlewycky@google.com">nlewycky@google.com</a>>, "clang-dev Developers" <<a href="mailto:cfe-dev@cs.uiuc.edu">cfe-dev@cs.uiuc.edu</a>><br>
> Sent: Friday, January 16, 2015 3:52:33 PM<br>
> Subject: Alias Analysis pass ordering in LLVM (and Clang)<br>
><br>
><br>
><br>
> (sorry for the wide distribution, but this impacts Clang users quite<br>
> a bit)<br>
><br>
> It's come up a few times in reviews of CFL-AA (a new alias analysis<br>
> for LLVM), so I wanted to write down what I think the current state<br>
> actually is for AA pass ordering, why, and how it should look<br>
> eventually. I may have some bugs here, so please correct me if I<br>
> miss anything. And I'd love thoughts about the end state.<br>
><br>
><br>
> Today, we have a strange ordering which is primarily motivated by<br>
> trying to support both TBAA and union-based local type punning:<br>
><br>
><br>
> 1) Run BasicAA. If it hits NoAlias, MustAlias, or PartialAlias, done.<br>
> If it hits MayAlias, delegate to #2.<br>
> 2) See if there is scoped no-alias metadata that can produce NoAlias<br>
> and return that if so. If not, delegate to #3.<br>
> 3) See if there is TBAA metadata that can produce NoAlias and return<br>
> that if so. If not, delegate to #4.<br>
> 4) If using CFL-AA, ask it and use its answer as the final answer.<br>
><br>
><br>
> The primary reason for running BasicAA first is so that it can return<br>
> MustAlias for local type puns, bypassing the more strict behavior of<br>
> TBAA. Having TBAA not fire on local type puns was a critical need<br>
> for the Darwin platform and maybe for others as well. However, this<br>
> implementation strategy is problematic for a host of reasons:<br>
><br>
><br>
> a) There are other users of LLVM that might want strict TBAA, and<br>
> they can't get that today.<br>
<br>
</div></div>They might, but because I can't imagine a use case where provably-incorrect answers are desired, it seems that (b) is really the underlying issue. That having been said, there might be a significant compile-time savings if TBAA, etc. is used extensively.<br>
<span class=""><br>
> b) BasicAA can hit partial-alias in cases that aren't even type puns<br>
> and which either TBAA or some other more-powerful AA like CFL-AA<br>
> would return no-alias. In these cases, we produce a too-conservative<br>
> answer.<br>
<br>
</span>I'd argue that these are bugs in BasicAA that should be fixed. BasicAA, like any other AA, should only return PartialAlias when it has *proved* a partial overlap. Otherwise, it should delegate. Papering over these by flipping the pass order is not desirable -- I know you're not proposing that, but I don't view bugs in BasicAA as a good motivation for a pass-ordering change.<br>
<span class=""><br>
> c) I'm worried that BasicAA does caching that isn't valid for all<br>
> possible AA implementations.<br>
><br>
<br>
</span>Interesting point; BasicAA's caching is supposed to serve as a guard against unbounded recursion, and is keyed off of the queried pair of AA::Location objects. I feel, however, that this must be universally valid. There is a complication, which is that Location's auxiliary metadata-based members (for TBAA, etc.) have control-flow dependencies, and so BasicAA can only cache results for locations with those members in blocks that are predecessors of the original query instructions (i.e. it can only walk up). I'm pretty sure this is what it does, but if not, there could be problems.<br>
<span class=""><br>
><br>
> I think we should change these things in the following way to get to<br>
> a more principled place.<br>
><br>
><br>
> First, we should make the order something a bit more obvious which<br>
> has to do with which AA passes have the most information:<br>
><br>
><br>
> 1) Run the scoped no-alias AA and return any final answers it<br>
> produces.<br>
> 2) Run TBAA and return any final answers it produces.<br>
> 3) Run CFL-AA (if enabled, or any other advanced AA implementations<br>
> we want) and return any final answers it produces (IE, anything more<br>
> precise than 'may-alias').<br>
> 4) Run BasicAA and use its answer.<br>
><br>
><br>
> I then think we should teach TBAA to have an explicit flag which<br>
> causes it, when discovering a 'no-alias' result, to query BasicAA<br>
> for a contradiction of must-alias or partial-alias, and to ignore<br>
> the no-alias in that case. That should more precisely model the<br>
> desired behavior of TBAA-except-for-local-type-puns.<br>
><br>
<br>
</span>I agree this would be cleaner.<br>
<span class=""><br>
><br>
> Finally, I think we should teach Clang to have two flags:<br>
> "-fstrict-aliasing" and '-fstrict-local-type-pun-aliasing'.<br>
<br>
</span>To engage in some bikeshedding, I'd prefer that we have:<br>
<br>
-fstrict-aliasing=<mode names>, where -fstrict-aliasing without a mode name has a target-dependent default, perhaps, as you've suggested. We could use a name like 'pedantic' or 'standard' for the stricter mode and 'loose' or 'heuristic' for the current mode.<br></blockquote><div><br></div><div><br></div><div>Adding flavor to -fstrict-aliasing seems reasonable.</div><div><br></div><div>David</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<span class=""><br>
>By<br>
> default, strict-aliasing can imply strict-local-type-pun-aliasing.<br>
> Darwin and any other platforms that want the current aliasing<br>
> behavior can make the use of strict-aliasing by default *not* imply<br>
> strict-local-type-pun-aliasing. This would still make it available<br>
> behind a flag, and let other platforms line up the defaults in other<br>
> ways that make sense to them.<br>
<br>
</span>I'm in favor of the additional conceptual clarity this would provide, and I also like the potential compile-time savings by not aggressively computing BasicAA results when TBAA, etc. can answer the query directly.<br>
<br>
Regardless, however, we should fix the exiting bugs in BasicAA's delegation/PartialAlias logic.<br>
<br>
-Hal<br>
<br>
><br>
><br>
> Thoughts?<br>
> -Chandler<br>
<span class="HOEnZb"><font color="#888888"><br>
--<br>
Hal Finkel<br>
Assistant Computational Scientist<br>
Leadership Computing Facility<br>
Argonne National Laboratory<br>
_______________________________________________<br>
cfe-dev mailing list<br>
<a href="mailto:cfe-dev@cs.uiuc.edu">cfe-dev@cs.uiuc.edu</a><br>
<a href="http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev" target="_blank">http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev</a><br>
</font></span></blockquote></div><br></div></div>