OCaml internals: Exceptions

Question

I'm curious to know how exceptions are dealt with in OCaml runtime to make them so lightweight. Do they use setjmp/longjmp or do they return a special value in each function, and propagate it?

It seems to me that longjmp would put a little strain on the system, but only when an exception is raised, while checking for each function return value would need to check for every and each value after calling a function, which seems to me would put a lot of checks and jumps, and it seems it would perform worst.

By looking at how OCaml interfaces with C ( http://caml.inria.fr/pub/docs/manual-ocaml/manual032.html#toc142 ), and looking at callback.h, it seems that an exception is tagged by using the memory alignment of objects ( #define Is_exception_result(v) (((v) & 3) == 2) ). This seems to indicate that its implementation doesn't use longjmp and checks each function result after each function call. Is that it? Or the C function already tries to catch any exception, and then converts it to this format?

Thank you!

Answer 1

OCaml exception handling

It doesn't use setjmp/longjmp. When a try <expr> with <handle> is evaluated, a "trap" is placed on the stack, that contains information about the handler. The address of the topmost trap is kept in a register¹, and when you raise, it jumps directly to this trap, unwinding several stack frames in one go (this is better than checking each return code). A trap also stores the address of the previous trap, which is restored in the register at raise time.

¹: or a global, on architectures with not enough registers

You can see for yourself in the code:

• bytecode compilation: lines 635-641, two Kpushtrap/Kpoptrap bytecodes surround the try..withed expression
• native compilation: lines 254-260, again instructions Lpushtrap/Lpoptrap around the expression
• bytecode execution for the bytecode PUSHTRAP (places the trap/handler), POPTRAP (remove it, non-error case) and RAISE (jump to the trap)
• native code emission on mips and on amd64 (for example)

Comparison with setjmp

Ocaml uses a non-standard calling convention with few or no callee-saved registers, which makes this (and tail-recursion) efficient. I suppose (but I'm no expert) that's the reason why C longjmp/setjmp isn't as efficient on most architectures. See for example this x86_64 setjmp implementation that looks exactly like the previous trapping mechanism plus callee-registers save.

This is taken into account in the C/OCaml interface: the usual way to call a Caml function from C code, caml_callback, doesn't catch OCaml-land exceptions; you have to use a specific caml_callback_exn if you wish to, which setups its trap handler and saves/restores callee-saved registers of the C calling convention. See eg. the amd64 code, which saves the registers then jump to this label to setup the exception trap.