* * * * *
The case of the regression test regression, part II
> When you have eliminated the impossible, whatever remains, however
> improbable, must be the truth.
>
> “Sherlock Holmes”
>
When last I left off [1] I identified the slow code in the regression test
and it left me puzzled—it was a single function call that did not change
between versions. Now, a bit of background: eight years ago [Eight
years⁈ Where did the time go? —Sean] [A world wide pandemic. —Editor]
[Gee, thanks. —Sean] I wrote a custom Lua interpreter [2] that contains all
possible Lua modules we could possibly use at work in order to avoid having a
bunch of code to install, which I call kslua (which stands for “Kitchen Sink
Lua”). And so far, that's what I've been using to run the regression test.
Faced with the fact that the sipsock:recv() call was taking upwards of a
second, I decided update just that module to the latest in the fast version
of the regression test as a sanity check. Well, it failed as a sanity check,
because the latest version of that module that contains that function ran
fast, so my sanity wasn't saved one bit. The only conclusion I can come to is
that something else has changed!
Fortunately, somethine else has changed. A bit more background: the
regression test is used to test “Project: Sippy-Cup [3],” “Project: Lumbergh
[4]” and “Project: Cleese [5].” And to run those programs, I need a few more
programs that those programs communicate with, and oh hey! There's a program
that the regression program runs that also runs via kslua! And through a
tedious process of elimination, I finally found a module that causes the
slowdown—the network event driver module [6] I wrote. I then went through a
tedious process of elminiation to find the exact change that causes the slow
down. The “fast” version of the function in question, which is written in C,
is:
-----[ C ]-----
static int polllua_insert(lua_State *L)
{
pollset__t *set = luaL_checkudata(L,1,TYPE_POLL);
int fh = luaL_checkinteger(L,2);
lua_settop(L,4);
if (set->idx == set->max)
/* ... */
-----[ END OF LINE ]-----
and the slow version, which is the next literal version of the code:
-----[ C ]-----
static int polllua_insert(lua_State *L)
{
pollset__t *set = luaL_checkudata(L,1,TYPE_POLL);
int fh;
lua_settop(L,4);
if (!luaL_callmeta(L,2,"_tofd"))
{
lua_pushinteger(L,EINVAL);
return 1;
}
fh = luaL_checkinteger(L,-1);
if (set->idx == set->max)
/* ... */
-----[ END OF LINE ]-----
I got tired of having to write (in Lua):
-----[ Lua ]-----
SOCKETS:insert(sock:_tofd(),'r',handler)
-----[ END OF LINE ]-----
so I changed the code to call _tofd() directly:
-----[ Lua ]-----
SOCKETS:insert(sock,'r',handler) -- the system will know to call _tofd()
-----[ END OF LINE ]-----
The only thing is—the program that calls this only calls this once in the
program.
At startup.
Desk, meet head.
So I'm again failing to see how this causes the slowdown. I use the “fast”
version and the regression runs fast. I click the version of that module one
step forward and it's slow.
It's maddening!
[1]
gopher://gopher.conman.org/0Phlog:2021/08/19.1
[2]
gopher://gopher.conman.org/0Phlog:2013/03/22.1
[3]
gopher://gopher.conman.org/0Phlog:2014/03/05.1
[4]
gopher://gopher.conman.org/0Phlog:2018/09/11.2
[5]
gopher://gopher.conman.org/0Phlog:2018/09/11.2
[6]
https://github.com/spc476/lua-conmanorg/blob/master/src/pollset.c
Email author at
[email protected]
