[clamav-users] Question about LLVM...
debian at kitterman.com
Tue Dec 11 13:30:12 EST 2018
On Tuesday, December 11, 2018 05:59:05 PM Micah Snyder wrote:
> Sorry about the broken links on the website and in the clamav-faq manual
> pages. Our web dev team is actively working on integrating the newly
> remodeled user manual into the website.
> The bytecode interpreter was nonfunctional for a long time but was fixed a
> few years ago. This is why LLVM was prioritized over the bytecode compiler.
> Functionally, from an outside perspective, the feature set of using bytecode
> interpreter vs LLVM is the same. The cost/benefit analysis of LLVM-JIT vs
> Interpreter hinges on whether or not executing native code is sufficiently
> faster than interpreting the bytecodes to outweigh the cost of JIT
> compilation. Our bytecode signatures themselves are relatively small and
> are relatively few, so the advantage of executing native code vs the time
> lost JIT compiling the bytecode is, I'm told, negligible. The developers
> who did the initial benchmarking on the subject have since left the team
> and while I've been told that the performance is "about the same", I don't
> have any figures to back up that up. If anyone out there decides to do
> additional research on the subject, do note that bytecode functions are
> only executed for certain file types, so benchmark findings will vary by
> file type.
> The TL;DR is that we're not aware of any significant advantage of using LLVM
> over the bytecode interpreter at this time.
> Regarding the reason for only supporting older versions of LLVM: It takes
> time to update to use newer APIs. The LLVM project has been moving pretty
> fast and we simply haven't prioritized dev and test time towards updating
> our LLVM support. In fact, Debian provides a patch to ClamAV to support
> LLVM 3.7-3.9, but we haven't had the time to properly integrate and test
> it. Because the bytecode interpreter is working so well, we're focusing
> our efforts on other tasks.
And unfortunately the developer who was doing that work in Debian has moved on
to other things, so we won't be providing patches for later versions.
Might it make sense in the next feature release to just kill off LLVM and move
on. That would certainly help with clarity and focus.
More information about the clamav-users