Late to the party because I figured other people would chime in, but I'll have a go...
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?
You need these to match so that running a pass using `opt` or `llc` on
a dumped bitcode file gets you the same result as running the pass
directly. This sort of thing (dumping out temporary results and
continuing from there) is useful when triaging bugs. The problem is
that the bug may not reproduce at all when starting from the dumped
bitcode.
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.
I'm not sure there's consensus that this is a bug. It's not optimal,
but it's not clear to me that it's invalid to depend on use-list order.
In some cases, it may be a reasonable heuristic that improves compile
time (I'm not arguing for that, I'm just not convinced otherwise).
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?)
(FWIW, there's some ancient discussion in PR5680 about this.)
I don't have a strong opinion on whether depending on use-list order
should be considered a bug. However, it *is* a bug not to be able to
roundtrip to bitcode and get the same end result.
While it may be possible to remove the compiler's dependency on use-list
order, no one has signed up to do the work, it's not clear what the
compile time cost would be, and there isn't consensus that it's the
right end goal.
In the meantime, this fixes the bug. If/when that hypothetical work is
done and validated we can turn this off by default if it makes sense to.
In terms of LTO bitcode size: serializing use-list order isn't actually
necessary/useful for the "normal" outputs of `clang -flto`. It's
important for `clang -emit-llvm`, `clang [-flto] -save-temps`, and
`<gold/libLTO> -save-temps`, but serialization between "compile" and
"link" is a deterministic and reproducible step. A possible
optimization would be to disable the option when writing `clang`'s
(final) output file in `clang -flto`. Thoughts?
Alternatively, some portions of the aforementioned hypothetical work
might be low-hanging fruit. E.g., it might not be hard to validate that
the use-list order of `ConstantInt`s doesn't affect output (and if it
does, it would probably be easy to fix); at the same time, I imagine
they account for a disproportionate percentage of the bitcode bloat
(they have large use-lists that change frequently).