The algorithm included in ObjectFileELF.cpp performs a byte at a time computation, which causes long pipeline stalls in modern processors. Unfortunately, the polynomial used is not the same one used by the SSE 4.2 instruction set, but there are two ways to make it faster:
Work on multiple bytes at a time, using multiple lookup tables. (see http://create.stephan-brumme.com/crc32/#slicing-by-8-overview)
Compute crcs over separate regions in parallel, then combine the results. (see http://stackoverflow.com/questions/23122312/crc-calculation-of-a-mostly-static-data-stream)
As it happens, zlib provides functions for both:
The zlib crc32 function uses the same polynomial as ObjectFileELF.cpp, and uses slicing-by-4 along with loop unrolling.
The zlib library provides crc32_combine.
I decided to just call out to the zlib library, since I see my version of lldb already links with zlib; however, the llvm CMakeLists.txt declares it optional.
I’m including my patch that assumes zlib is always linked in. Let me know if you prefer:
I make the change conditional on having zlib (i.e. fall back to the old code if zlib is not present)
I copy all the code from zlib and put it in ObjectFileELF.cpp. However, I’m going to guess that requires updating some documentation to include zlib’s copyright notice.
This brings startup time on my machine / my binary from 50 seconds down to 32.
(time ~/llvm/build/bin/lldb -b -o ‘b main’ -o ‘run’ MY_PROGRAM)
zlib_crc.patch (6.17 KB)
Zlib is definitely optional and we cannot make it required.
Did you check to see if llvm has a crc32 function somewhere in Support?
I didn’t realize that existed; I just checked and it looks like there’s JamCRC which uses the same polynomial. I don’t know what “Jam” means in this context, unless it identifies the polynomial some how? The code is also byte-at-a-time.
Would you prefer I use JamCRC support code instead, and then change JamCRC to optionally use zlib if it’s available?
It would be nice if we could simply update LLVM’s implementation to be faster. Having multiple implementations of the same thing seems undesirable, especially if one (fast) implementation is always superior to some other reason. i.e. there’s no reason anyone would ever want to use a slow implementation if a fast one is available.
Can we change the JamCRC implementation in LLVM to use 4-byte slicing and parallelize it ourselves? This way there’s no dependency on zlib, so even people who have non-zlib enabled builds of LLDB get the benefits of the fast algorithm.
BTW, the JamCRC is used in writing Windows COFF object files, PGO instrumentation, and PDB Debug Info reading / writing, so any work we do to make it faster will benefit many parts of the toolchain.
What about the crc combining? I don’t feel comfortable reimplementing that on my own. Can I leave that as a feature predicated on zlib?
For the JamCRC improvements, I assume I submit that to llvm-dev@ instead?
It seems like the the crc32_combine is not too hard to implement, but we do already have code in LLVM that is predicated on the existence of zlib, so it seems reasonable to leave it that way. And yes, you would submit those changes to llvm-dev. Perhaps the JamCRC implementation could be updated to have the conditional logic already embedded, so that the only implementation is the JamCRC implementation, and people don’t have to think about how to get the “fast” one.
Ok I stripped out the zlib crc algorithm and just left the parallelism + calls to zlib’s crc32_combine, but only if we are actually linking with zlib. I left those calls here (rather than folding them info JamCRC) because I’m taking advantage of TaskRunner to parallelize the work.
I moved the system include block after the llvm includes, both because I had to (to use the config #defines), and because it fit the published coding convention.
By itself, it reduces my test time from 55 to 47 seconds. (The original time is slower than before because I pulled the latest code, guess there’s another slowdown to fix).
zlib_crc.patch (8.96 KB)
I know this is outside of your initial goal, but it would be really great if JamCRC be updated in llvm to be parallel. I see that you’re making use of TaskRunner for the parallelism, but that looks pretty generic, so perhaps that could be raised into llvm as well if it helps.
Not trying to throw extra work on you, but it seems like a really good general purpose improvement and it would be a shame if only lldb can benefit from it.
Improving the checksumming speed is definitely a worthwhile contribution, but be aware that there is a pretty simple way to avoid computing the crc altogether, and that is to make sure your binaries have a build ID. This is generally as simple as adding -Wl,–build-id to your compiler flags.
+1 to moving the checksumming code to llvm
There is a good crc32c (assuming we want crc32c) code in DPDK
It has hardware assisted algorithm for x86 and arm64 (if hardware
supports it). There is a fallback to lookup table implementation.
CRC32 is definitely worth merging with LLVM.
Interesting. That saves lldb startup time (after crc improvements/parallelization) by about 1.25 seconds wall clock / 10 seconds cpu time, but increases linking by about 2 seconds of cpu time (and an inconsistent amount of wall clock time). That’s only a good tradeoff if you run the debugger a lot.
If all you need is a unique id, there are cheaper ways of going about it. The SSE crc instruction would be cheaper, or using CityHash/MurmurHash for other cpus. I thought it was specifically tied to that crc algorithm. In that case it doesn’t make sense to fold this into JamCRC, since that’s tied to a difficult-to-optimize algorithm.
What we need is the ability to connect a stripped version of an SO to one with debug symbols present. Currently there are (at least) two ways to achieve that:
build-id: both SOs have a build-id section with the same value. Normally, that’s added by a linker in the final link, and subsequent strip steps do not remove it. Normally the build-id is some sort of a hash of the initial file contents, which is why you feel like you are trading debugger startup time for link time. However, that is not a requirement, as the exact checksumming algorithm does not matter here. A random byte sequence would do just fine, which is what “–build-id=uuid” does and it should have no impact on your link time. Be sure not to use this if you care about deterministic builds though.
gnu_debuglink: here, the stripped SO contains a checksum of the original SO, which is added at strip time. This is done using a fixed algorithm, and this is important as the debugger needs to arrive at the same checksum as the strip tool. Also worth noting is that this mechanism embeds the path of the original SO into the stripped one, whereas the first one leaves the search task up to the debugger. This may be a plus or a minus, depending on your use case.
Hope that makes things a bit clearer. Cheers,
Thank you for that clarification. Sounds like we can’t change the crc code then.
I realized I had been using GNU’s gold linker. I switched to linking with lld(-4.0) and now linking uses less than 1/3rd the cpu. It seems that the default hashing (fast == xxHash) is faster than whatever gold was using. I’ll just switch to that and call it a day.
FreeBSD's default toolchain still uses an ancient GNU ld that lacks
build-id support (on all platforms other than aarch64), so I'll be
very happy to see improvements in checksum speed. We're migrating to
LLD and will be able to use --build-id eventually, but it will be a