libera/#sicl - IRC Chatlog

3:57:06 beach I was thinking yesterday about how literals are accessed from native code. Maclina uses a vector of literals, and I guess I hadn't considered that possibility, though that might be what mainstream Common Lisp implementations do for native code.

3:58:31 beach But that would be quite expensive for an ordinary function call. Whether you use the symbol as an indirection as I suppose most Common Lisp implementations do, or whether you use the "cell" indirection as we do with Clostrum, the symbol or the cell would be considered a literal that must be accessed.

3:59:07 beach So if you have a vector, then that vector needs to be passed around to every function call, so it has to be accessed from the function object.

3:59:31 beach Then, when a function call is to be made, an entry of that vector must be accessed.

4:00:02 beach That's two additional memory operations for each function call.

4:01:07 beach In the past, I considered passing such literals in the closure vector, but that's even worse, because then they have to be passed to intermediate functions that don't use them.

4:02:12 beach But then I decided for SICL that I would make literals have fixed addresses, so they could be loaded from the instruction stream using one or two native instructions.

4:04:01 beach So right then I was thinking that with call-site optimization, the issue disappears, or rather, it is replaced by the need to patch native code to insert a different address for the trampoline snippet.

4:04:38 beach But that won't fix literals other than function cells.

4:05:47 bike in native clasp, the literals vector (which isn't an actual lisp vector) is stored close to the actual code, and then when the code grabs a literal it just loads it from an address relative to the instruction pointer

4:06:07 beach Ah, nice!

4:06:28 beach Is that what SBCL is doing as well?

4:08:14 beach A PC-relative load is probably as inexpensive as a constant load in modern processors.

4:08:36 bike if i disassemble (lambda () 'x) i see ; 54: 488B15C5FFFFFF MOV RDX, [RIP-59] ; 'X

4:08:39 bike so yeah, i think it is

4:08:59 beach That's good to know. Thanks!

4:10:43 beach I suppose you could even have the literals move if you allow the garbage collector to access that part of the instruction space and modify it.

4:14:13 bike yeah, and storing the literals and code together means they can be collected together easily, too. not that we do that in clasp yet

4:14:32 beach Yes, good point.