Re: start-up time

[Andrew Haley <aph at redhat dot com>]

Per Bothner writes:
 > A little experiment, running and timing this command:
 >    $KAWA -e '(define (hello s) (display s) (newline)) (hello "Hello!")'
 > This is basically "hello world", but it also compiles a function and
 > then calls it, which requires a fair bit of Kawa to get started.
 > 
 > With different values of $KAWA:
 > 
 > (1) java version "1.4.2_11"
 > Java(TM) 2 Runtime Environment, Standard Edition (build 1.4.2_11-b06)
 > Java HotSpot(TM) Client VM (build 1.4.2_11-b06, mixed mode)
 > 
 > Time: about 1s, whether using a .jar or .class files.
 > 
 > java version "1.5.0_05" is about the same.
 > 
 > (2) gij (GNU libgcj) version 4.1.1 20061011 (Red Hat 4.1.1-30)
 > 
 > Time: actually slightly faster, about 0.9s.
 > 
 > (3) gcj-compiled Kawa, with -findirect-dispatch, and installed into a
 > temporary location (i.e. without libtool indirection).
 > 
 > Time: about 5s.  Ouch.
 > 
 > (4) /usr/bin/kawa (older version of Kawa) installed from the Fedora
 > Extras rpm:
 > 
 > Time: about 0.45s.
 > 
 > (5) gcj-compiled kawa, then running running sudo gcj-dbtool,
 > followed by gij -jar
 > 
 > Time: about 0.5s.  (The difference between (4) and (5) is probably
 > not significant, and could be because of different Kawa versions.)
 > 
 > So gcj-dbtool seems to be dong some extra magic besides just entering
 > the jar/so in a database.  Some kind of pre-linking?

No.  gcj-dbtool adds a file to a database, that's all.  I have no
explanation for (3).

 > Anyway, I thought this was instructive.
 > 
 > I'm thinking the right thing for Kawa's 'make install' is after
 > copying the .jar and .so to their target locations, try to run
 > gcj-dbtool.  That means checking that gcj-dbtool exists and also that
 > the database is writable (i.e.: test -w `gcj-dbtool -p`).
 >
 > Does that sound reasonable?

That should work.

 > Or does the Fedora aot-compile script do anything else I should do?

No, it doesn't.

Andrew.