Does PolkaVM do static or dynamic binary translation?

[ghost]

On first look PolkaVM seems to use a static translation approach to me, so I was curious what constraints you have on the original ELF. If I'm wrong and it's a dynamic approach, could you provide a rough overview of the translation strategy?

[koute]

It's a single pass recompiler that translates the whole module when it's instantiated.

Here's our toolchain pipeline:

Basically the idea is to move as much code and complexity offline into polkavm-linker. The polkavm-linker is what actually ingests a vanilla RISC-V ELF file and translates it into a form that the VM/recompiler can accept. It supports arbitrary programs as long as no self-modifying code is present and the program doesn't try to inspect its own raw RISC-V machine code nor has any dynamic jumps which depend on the RISC-V machine code offsets being preserved (so 99.9% of programs are compatible).

[ghost]

Makes sense. How about programs that interleave code and data? I'm guessing that's also not allowed?

[koute]

If they're in the same section then currently that won't work (although that could probably be somewhat supported if we'd waste space potentially having to include the original RISC-V machine code in the blob). Fortunately in practice compilers don't emit such programs; modern toolchains essentially always emit code and data in their own separate sections.