Suggest using preserve_none to improve tail call optimization in interpreter

[yangyi625]

I’ve been exploring rv32emu's interpreter mode and found that while the RVOP macro is designed to support MUST_TAIL tail calls via:

MUST_TAIL return next->impl(rv, next, cycle, PC);

In practice, the compiler often emits full stack frames and callee-saved register spills, which breaks the intended tail call optimization (TCO). This happens even though the control flow would otherwise be eligible for a proper tail jump.

For example, in a simple handler like do_addi, the compiler emits a proper tail call via jmp *%rax, and avoids saving any callee-saved registers:

af38:   ff e0    jmp *%rax   # tail call to next->impl

However, in do_lw, due to a non-tail call (e.g., memory access through a function pointer), the compiler emits callee-saved register preservation and a full frame:

aaa0:   55        push   %rbp
aaa1:   41 57     push   %r15
...
ab75:   c3        ret

This overhead defeats the benefits of using MUST_TAIL, especially in a hot loop of instruction dispatch.

##Suggestion

Clang provides the attribute((preserve_none)) attribute (available since 19.1.0) which modifies the calling convention to avoid saving any registers. It’s suitable when the function is only ever tail-called and doesn’t return via ret.

This attribute allows better register usage and completely avoids unnecessary stack setup for tail calls — perfect for interpreter dispatch functions.

References

Clang docs: preserve_none

A Tail Calling Interpreter For Python (And Other Updates)