On 2020-05-18 07:58, Magnus Ihse Bursie wrote:
On 2020-05-18 15:16, Erik Joelsson wrote:
Should we then add the big flags instead to have some control over
memory footprint, or are we trusting the default ergonomics to do the
right thing on every possible build system out there?
You make it sound like we shouldn't. :-) But seriously, I think we
should have some faith in our product to actually produce good default
values. If anything, the "small" flags were created to signal that the
default ergonomics didn't really apply, that we favored fastest
possible startup, at the expense of long-term performance.
Well we are making some attempt at creating a suitable set of BIG flags
in configure based on the build system, but that is mainly to make sure
we allocate enough resources to the javac server. On second thought I
agree that the main reason for java flags in the build is to keep the
footprint of most of them small. Unless there was a measurable
performance increase using BIG for jlink, there is no point.
Also, this is what we do for all other tools. There is no
"$(JAVA_TOOL_FLAGS_BIG)". Yes, I know, there is a $(JAVA_FLAGS_BIG),
and we could of course turn it into a "TOOLS" version, but so far we
have not done that. (The difference is that the "normal" version is
supposed to be passed to java, like -Xmx, and the "TOOLS" version is
supposed to be passed to tools like javac, so they are prefixed with
-J like -J-Xmx.)
Are you happy with this, or do you want me to change the patch to
include a new JAVA_TOOL_FLAGS_BIG?
Nah, I'm ok with it. Looks good.
/Erik
/Magnus
/Erik
On 2020-05-18 03:12, Magnus Ihse Bursie wrote:
The linking of the JDK image using the jlink tool takes a
considerable amount of time. However, the jlink tool is classified
as a "small java" tool in the build script. This should only be used
on quick small tools where a long startup time has a measurable
performance impact. Clearly this does not apply to jlink.
The effect of removing the small tool flags are most pronounced when
running a slowdebug build. Robbin Ehn, which made me aware of this,
reduced the running time of jlink from 36 to 21 seconds on his
machine by removing the "small" flags. On a release build, the
effect is much less noticeable, but it's still worth fixing.
Bug: https://bugs.openjdk.java.net/browse/JDK-8245168
Patch inline:
diff --git a/make/autoconf/spec.gmk.in b/make/autoconf/spec.gmk.in
--- a/make/autoconf/spec.gmk.in
+++ b/make/autoconf/spec.gmk.in
@@ -642,7 +642,7 @@
JAVAC=@FIXPATH@ $(JAVAC_CMD)
JAVADOC=@FIXPATH@ $(JAVADOC_CMD)
JAR=@FIXPATH@ $(JAR_CMD)
-JLINK = @FIXPATH@ $(JLINK_CMD) $(JAVA_TOOL_FLAGS_SMALL)
+JLINK = @FIXPATH@ $(JLINK_CMD)
JMOD = @FIXPATH@ $(JMOD_CMD) $(JAVA_TOOL_FLAGS_SMALL)
JARSIGNER=@FIXPATH@ $(JARSIGNER_CMD)
JJS=@FIXPATH@ $(JJS_CMD) $(JAVA_TOOL_FLAGS_SMALL)
/Magnus
