On Thu, May 7, 2020 at 11:09 AM John Fastabend <john.fastabend@xxxxxxxxx> wrote: > > I think it would work but I it would be even nicer if clang, verifier > and jit caught the tail call pattern and did it automatically. I've been advocating for proper tail calls for some time :) All it needs is indirect jump instruction in the ISA. The changes to llvm are trivial. Encoding of new insn is straightforward as well. The verifier side is tricky. What you're proposing makes sense to me. Somehow I thought that we need full indirect jump from day one, but above is much simpler. It's a subset of it. It's still an indirect jump, but target is always fixed. The register will be initialized with fixed address of next kernel function (or helper). That should be easy enough to support in the verifier. llvm will generate: ld_imm64 rX = addr_of_next_helper // that could be encoded via pseudo, like for calls to helpers jmp *rX We can introduce an extension to JA insn instead that takes 64-bit immediate or pc relative offset, but I think it will be more messy to support through llvm, libbpf relocations and the verifier.
