The x86_64 qvm JIT compiler is very slow, often taking around 5 seconds to compile a qvm. This patch caches compiled vms so that a recompile is not necessary when the same qvm is reloaded. I also disabled a debug #ifdef which causes unnecessary instructions to be emitted.
We should profile the code and see why it's slower than the other JIT compilers.
A cache might be interesting at the higher level, so all JITs can use it, but if they're basically all fast except for one, and we can optimize that one, we probably don't need the cache at all.
--ryan.
Um, I don't even have to profile this. The x86_64 JIT is slow because it writes out assembly and spawns gas to assemble it.
That should probably be fixed.
--ryan.
(In reply to comment #4)
> Um, I don't even have to profile this. The x86_64 JIT is slow because it writes
> out assembly and spawns gas to assemble it.
It has it's own assembler, gas was only used temporarily. With that
intermediate assembler step it's still slower than hardcoding
opcodes as in vm_86 of course.
I'm sorry, but this VM caching stuff is not going to happen. A user could overwrite previously cached VMs much more easily and thus circumvent the pure checks more easily.
However, if you've got a patch which would remove the intermediary assembling step and thus improve speed I'd be willing to look at this again.
Created attachment 2085 [details] The patchCreated attachment 2086 [details] Updated patch that fixes some bugs