[LLVMdev] [RFC] Setting preserve-bc-use-list-order=true by default

[David Blaikie]

Late to the party because I figured other people would chime in, but I'll
have a go...

On Tue, Mar 31, 2015 at 7:10 PM, Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:

> A while back I finished up some work [1] that Chad started to preserve
> use-list-order in bitcode [2], hidden behind an "experimental" option
> called `-preserve-bc-use-list-order`.  I then added a similar
> `-preserve-ll-use-list-order` option for LLVM assembly [3].
>
> [1]: http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-July/074604.html
> [2]: https://llvm.org/bugs/show_bug.cgi?id=5680
> [3]: https://llvm.org/bugs/show_bug.cgi?id=20515
>
> I'd like to move both of these options out of "experimental" mode, and
> turn `-preserve-bc-use-list-order` on by default.  I've attached a patch
> that does this.
>
> Why?
> ====
>
> Use-list order affects the output of some LLVM passes.  A typical
> example is a pass that walks through basic block predecessors.  The
> use-list order is deterministic, but not preserved when serializing the
> LLVM IR to bitcode or assembly.
>
> In rare (but frustrating) cases, this makes it difficult to reproduce a
> crash or miscompile from an intermediate result.
>
> For example, consider running an LTO build and serializing between the
> LTO optimization pipeline and the LTO codegen pipeline.  On SPEC,
> serializing to/from bitcode will change the output executable in 33
> benchmarks.  If you use `-preserve-bc-use-list-order`, all executables
> match.
>

Why do you need those to match? What's the penalty/problem when these don't
match?

Reproducibility for tests/bugs/etc seems important, but if it's any
worse/better with a particular use list, that's a bug we should fix, right?
Forcing a specific use list isn't going to fix those bugs, just make sure
we make the same decision (good or bad) every time.


> How?
> ====
>
> Have a look at `BitcodeReader::ParseUseLists()` and
> `predictUseListOrder()` in `lib/Bitcode/Writer/ValueEnumerator.cpp`.
>
> Does it degrade smoothly if IR changes?
> =======================================
>
> Yes.  If a `Value`'s use-list has changed in size (e.g., because of
> bitcode upgrades), its use-list order record is ignored.
>
> What does it cost?
> ==================
>
> Manman collected a bunch of numbers, with `-preserve-bc-use-list-order`
> and `-preserve-ll-use-list-order` both turned on:
>
>   - Time increase on LTO build time: negligible.
>   - Filesize increase in bitcode files output by `clang -flto`:
>       - clang: 6.8556% (sum when from 310412640 to 331693296 bytes).
>   - Filesize increase in merged bitcode files (output of
>     `ld -save-temps` when using `clang -flto` on Darwin).
>       — SPEC: 5.24% to 23.4% increase in file size.
>       — clang and related tools: 0.83% to 9.19% (details in
>         filesize_clang.txt, attached).
>   - Time increase of running `llvm-dis` on merged bitcode files:
>       — 6.4% to 15% on SPEC benchmarks with running time >= 1 second.
>       — 7.9% to 14.5% on clang with running time >= 1 second (details in
>         dis_clang.txt, attached).
>   - Time increase of running `llvm-as` on merged bitcode files:
>       — 3.5% to 39% on SPEC benchmarks with running time >= 1 second.
>       — 14.7% to 24.5% with running time >= 1 second (details in
>         as_clang.txt, attached).
>

These seem like pretty big costs to pay (bitcode size is going to be
particularly important to Google - big projects under LTO, limits on the
total size of the inputs to the link step, etc). To the point above, it's
not clear why we'd want to pay that cost. (I'm partly playing devil's
advocate here - I realize reproducibility is really important for a bunch
of reasons, though this particular reproducibility is a marginal one
(compared to "run the compiler twice on the same input and get two
different answers") but it seems like we've generally treated these issues
as bugs and fixed the optimizations to be use-list-order independent in the
past, no?)


>
> Does it work?
> =============
>
> Manman also ran `verify-uselistorder` (a use-list fuzzer that shuffles
> use-lists, serializes IR, compares use-lists, rinses, repeats), on the
> merged bitcode files above.  After I fixed a bug in assembly
> serialization in r232200, they all come up clean.
>
> Why leave -preserve-ll-use-list-order=false by default?
> =======================================================
>
> Firstly, our practise is to error on mismatches in LLVM assembly (unlike
> bitcode where we upgrade or drop old formats).  Printing use-list
> directives by default would make it almost impossible to update an
> assembly file by hand without stripping out the directives.
>
> Secondly, use-list order is encoded as a "shuffle", which isn't
> human-friendly.  I'm not convinced this adds value to textual IR except
> in rare cases where a developer actually wants to see the order, or
> wants to try to make a change without perturbing the use-lists.
>
> Nevertheless, the option works.  An LLVM developer can turn it on and
> hack at the IR, and `llvm-as` will complain if the developer makes the
> `uselistorder` directives inconsistent.  The developer can choose
> whether to delete the mismatched directives or to update them.
>
>
>
> _______________________________________________
> 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/20150409/e03ee67e/attachment.html>