On 8/12/13 8:41 AM, Alan Bateman wrote:
On 12/08/2013 14:26, harold seigel wrote:
The change to the MakeClasslist tool to use TraceClassLoadingPreorder
was made prior to my changes. The following code has been in
MakeClasslist.java for a long time. I was unable to determine who had
made this change or why.
// Understand only "Loading" from
-XX:+TraceClassLoadingPreorder.
// This ignores old "Loaded" from -verbose:class to force
correct
// classlist generation on Mustang.
if (t.equals("Loading")) {
t = tok.nextToken();
t = t.replace('.', '/');
So, I don't know why TraceClassLoadingPreorder was chosen instead of
-verbose:class or TraceClassLoading.
The purpose of my changes are to update the comments to accurately
reflect what the code currently does and to fix a bug in
MakeClasslist.java's 'for' loop where the wrong set of strings is used.
Okay, I see this was changed to use TraceClassLoadingPreorder in 2006
(prior to OpenJDK). Brent might know more about the history.
I know that for a long while, -verbose:class output was used to generate
the classlist. I did some digging...
In JDK 6, we started using TraceClassLoadingPreorder because it helped
line things up better for cold start [1] - better order of disk reads,
cut down on seeks, etc. I myself am not familiar with the details of
how this was made to work.
There was a short-lived version of MakeClasslist that understood output
from both -verbose:class and -XX:+TraceClassLoadingPreorder. The
nightly build scripts at the time were used for building both JDK 5 &
JDK 6 trains, so it had to work with both.
Later, support for -verbose:class was removed to ensure that the
classlist for 6+ was always created based on TraceClassLoadingPreorder,
to be sure we'd get the cold start benefit.
So that's what the comment about ignoring -verbose:class output is about.
-Brent
[1] http://bugs.sun.com/view_bug.do?bug_id=6449635 : "Optimize shared
archive object ordering for better cold startup performance"
