search

wahern / cqueues

Continuation Queues: Embeddable asynchronous networking, threading, and notification framework for Lua on Unix.
http://25thandclement.com/~william/projects/cqueues.html
MIT License
244 stars 37 forks source link

Use alignment hack to workaround luajit lightuserdata #243

Closed daurnimator closed 3 years ago

daurnimator commented 4 years ago

Idea from https://github.com/wahern/cqueues/issues/241#issuecomment-669661708

@vcunat please review?

vcunat commented 4 years ago

:thinking: cqueues seem to be using that mask on various static const int (defined inside functions), and there I don't see any guarantee that they will be placed at least 32 bytes apart.

I also don't know if the mentioned "negative pointers" are realistic; they would cause problems

0xFFF0000000000000 to 0xFFFFFFFFFFFFFFFF

– not just for uniqueness but they would still be over 47 bits in the code as written ATM.

daurnimator commented 4 years ago

thinking cqueues seem to be using that mask on various static const int (defined inside functions), and there I don't see any guarantee that they will be placed at least 32 bytes apart.

Ah yep... I wonder if results from lua_newuserdata are guaranteed to fit in a lightuserdata. But looks like we'll need to different methods.

I also don't know if the mentioned "negative pointers" are realistic; they would cause problems

0xFFF0000000000000 to 0xFFFFFFFFFFFFFFFF

– not just for uniqueness but they would still be over 47 bits in the code as written ATM.

ah right; we still need the mask.

Another thought: we should mix the high bits into the low bits to preserve uniqueness if we use multiple methods.

vcunat commented 4 years ago

Uniqueness: if you shift by six bits and use a 47-bit mask, you'll get the "sign" bit in the most significant one. EDIT: so possibilities are there, but the overall space to cover might be a bit large.

daurnimator commented 4 years ago

re: uniqueness I was thinking of if you ended up with e.g.

0x0000001234567800 as cqueue__poll (imagining we shifted by 8) and 0x0000000012345678 as something returned from lua_newuserdata

then they could collide. so instead of shifting 0x0000001234567800 right by 8; we'd move the high bits to low: x | ((x & 52) >> (52 - 8)) or something

vcunat commented 4 years ago

I didn't realize you want to avoid collisions with other values, too.

vcunat commented 3 years ago

So... we give this up in favor of https://github.com/LuaJIT/LuaJIT/commit/e9af1abec54 ?

daurnimator commented 3 years ago

So... we give this up in favor of LuaJIT/LuaJIT@e9af1ab ?

ooo, yeah that sounds like a reasonable idea.... does it all work well?

vcunat commented 3 years ago

I haven't tried it yet, but on a quick look, the approach supports up to 256 37-bit segments, so that sounds like a nice compromise.

daurnimator commented 3 years ago

Closing as https://github.com/LuaJIT/LuaJIT/commit/e9af1abec54 is the real fix.

We should remove the current 47bit hack code in the next release.