Follow-up of #195
We should be able to get throughput in AOT binaries to a very similar point as in JIT, if we provide PGOs to the compiler. We are currently using ML-inferred PGOs whoch are better than nothing, but can't possibly know what are the priorities for jelly-cli. If we had actual PGO info, we could probably both speed up Jelly-related code, and reduce the size of the code that we care less about (I'm looking at you, Xerces).
A good load would use all available formats, all Jelly features, and all commands. This will be pretty laborious... so it's not a huge priority for the moment.
Follow-up of #195
We should be able to get throughput in AOT binaries to a very similar point as in JIT, if we provide PGOs to the compiler. We are currently using ML-inferred PGOs whoch are better than nothing, but can't possibly know what are the priorities for jelly-cli. If we had actual PGO info, we could probably both speed up Jelly-related code, and reduce the size of the code that we care less about (I'm looking at you, Xerces).
A good load would use all available formats, all Jelly features, and all commands. This will be pretty laborious... so it's not a huge priority for the moment.