45300+ entries in 0.027s
mircea_popescu: in no case do i know of anyone who has actual data re such
things as "ok, so manual claims, but NUMBERS for
this penalty"
mircea_popescu: nevertheless,
the point has legs -- what we've done here very much can be done ands re-done, and with cuntoo/ada-gnat/etc stack spitting out statics, it might even come close.
mircea_popescu: kinda what i understood,
that cmp USED
TO BE expensive, but is no longer.
mircea_popescu: if
the zcx's cmp WERE slower
than sjlj's
test,
then we should see
the latter be faster on 0 handlers
than
the former!
mircea_popescu: she corrected
the numbers, it's 90 for zcx 93 for sjlj at
the best.
mircea_popescu: well, i won't
trust my own understanding of asm and contemporary cpus as far as i can
throw it ; but if indeed
the operands in zcx impl were slower, you'd see it
take less
time!!!1
mircea_popescu: yeah but
the pipe is built such
that
this is also ~0 cost electrically
mircea_popescu: one part of
the problem might be
that sjlj comes from a
time before, when insanities like
that snippet above were standard. but no
time since
the millenium do you see it instead of
the cmp etc.
mircea_popescu: are you basically saying
this sub eax, 1 ;
test eax, eax ; jz loc_601 is optimal approach ?
mircea_popescu: one
that's due
to
the method, and
the other
that's due
to
the fact zcx was a lot narrowly-er massaged
mircea_popescu: rather, i can't shake
this impression
that sjlj saddles us with
two segments of overhead
diana_coman: I'm atm doing
the inventory of ave1's versions of gnat scripts and apparently even 2018-05-29 relies on downloading stuff
that meanwhile moved/vanished as
they always do; moreover, I have
the darned stuff , now need
to figure out how
to cut out
the download and just point
the script at local source, ugh
mircea_popescu: (the 52 is cuz i
took
the 13 items and multiplied by 4, forgetting
that
these are actually byte alligned not 64-bit alligned)
mircea_popescu: what does it do with
the rest of
the frame, from
the bytes we see
to
the 184 ?
mircea_popescu: no i know where you got
the sub param from, what im asking is,
mircea_popescu: the observation
that perhaps sjlj is not actually as
tightly optimized as zcx is
trying
to percolate
through my brain
mircea_popescu: look here : lines 1
through 9 in zcx add up
to 13 bytes, yes ?
diana_coman: asciilifeform,
the name of
the
tarball is correct; you'll have
to change
the name of
the dir /put
them separate
diana_coman: asciilifeform, here are
the latest aka proc calls with 3 handlers per proc: ossasepia.com/available_resources/bins_calls_sjlj_adacoregnat.tar.gz and ossasepia.com/available_resources/bins_calls_zcx_adacoregnat.tar.gz ; let me know if you want anything else
diana_coman: asciilifeform, any preference re "pair of bins" i.e.
the procedure calls or
the loops of yest?
diana_coman: in other
things, re
http://btcbase.org/log/2019-02-11#1894896 ->
this should now be fully sorted i.e. IP change for dianacoman.com propagated as far as I can see + redirection working fine for any link so please let me know if you still encounter
trouble with any dead links; if you use only hosts (no DNS)
then simply adding dianacoman.com on same IP as ossasepia should work seamlessly
☝︎ diana_coman: ah,
that'd explain it, wouldn't it: by
the
time "programming" is direct
translation of fuzzing into code, it'd possibly speed up, yes
mircea_popescu: prolly a bunch of
try()catch semantics in "all programs"
diana_coman: I still don't see
the boo-boo of docs i.e ~"all programs should see a great improvement running zcx" or how was it
diana_coman: sorry; I should know by now
to not hurry up with data report even if it's just 2 runs
diana_coman: all
those
tests are on Adacore's 2016 gnat, yes
diana_coman: no, fat-fingered it, 0 instead of 9 i.e it was 0.93 not 0.03; sorry about
that; still running atm
the 1 handler with sjlj and
then will move on
to 2 and 3 handlers
mircea_popescu: ^even handled sjlj is not really
that bad, 7us per call far far from end of world.
mircea_popescu: (in fairness
though, no program ever does
the sort of calling insanity we do here, so irl
this may be very mild indeed)
mircea_popescu: it really blew my fucking mind! ZERO COST,
they said!!!
diana_coman goes
to run and will be back with proper data
diana_coman: and yes,
then I'll do with 2 and 3 handlers
too
diana_coman: ugh, either I fat-fingered
there or what; let me run
that again ; (and possibly /me should really stop getting data *other*
than in a nice plain
table)
mircea_popescu: ie, what
the docs don't say is
the juciest bit at all : if you do not have extra handlers, zcx is MASSACRING you on calls.
mircea_popescu: specifically stated,
this program
takes
to run : 1 with sjlh, no handlers ; 30 (up 3000%) with zcx, irrespective of handler count ; 5295 (a further 200% up) with sjlj and one extra handler.