NYI bytecode and some questions
- From: John Abrahamsen <jhnabrhmsn@xxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Tue, 18 Jun 2013 20:49:32 +0200
Hi all.
I'm doing a http(s)/web framework completely in Lua. All syscalls are
done directly via the FFI, this includes socket and epoll handling.
SSL support is also in there :).
And its actually coming along nicely, obviously faster than Tornado on Pypy.
I've lately been running LuaJIT with -jdump.
What I keep seeing obviously when benchmarking the framework itself
(no processing in requests except push "Hello world" to socket) is
that the traces are aborted frequently because of NYI bytecode. What I
see most of the time is:
0004 . UGET 3 0 ; ffi
0005 . TGETS 3 3 0 ; "new"
0006 . KSTR 4 1 ; "struct epoll_event[?]"
0007 . MOV 5 2
0008 . CALL 3 2 3
0000 . . FUNCC ; ffi.new
---- TRACE 6 abort epoll_ffi.lua:144 -- NYI: unsupported C type conversion
This is basically a trace starting at the top of the event loop and
break at the end when the array for epoll_wait is allocated. Another
one that keeps popping up is bc 51:
0008 . . . TGETS 3 0 2 ; "io_loop"
0009 . . . MOV 4 3
0010 . . . TGETS 3 3 3 ; "add_callback"
0011 . . . FNEW 5 4 ; iostream.lua:353
---- TRACE 9 abort iostream.lua:363 -- NYI: bytecode 51
This is from creating lambda functions.
My questions are; should I care if they are not implemented?
Should I rewrite the framework to avoid lambdas? The API is mostly
callbacks based at the moment for convinience. Is there a way to avoid
the two aborts mentioned?
Otherwise a great piece of engineering and many thanks to Mike!
/John
Other related posts:
