Andrew John Hughes wrote:
2009/10/15 Tim Bell <tim.b...@sun.com>:
Andrew John Hughes wrote:
The fix for bug 6888888:
http://hg.openjdk.java.net/jdk7/build/jdk/rev/777714bd992a
breaks the build as it tries to use javah from ALT_JDK_IMPORT_DIR (via
JAVA_TOOLS_DIR) which is not set on normal builds.
The README tells the user to set ALT_BOOTDIR and changing
JAVA_TOOLS_DIR to be set to BOOTDIR makes the build work again:
Sorry about that - my mistake.
As Andrew pointed out on IRC:
you can work around the bug by setting ALT_JDK_IMPORT_DIR to the same as
ALT_BOOTDIR
Jon Gibbons has a fix to 6889255 out for review, which will remove the 6888888
workaround:
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=6889255
http://cr.openjdk.java.net/~jjg/6889255-classreader.1
http://cr.openjdk.java.net/~andrew/javah/webrev.01/
Ok to push?
Unfortunately it is too late to fix in b74, and hopefully 6889255 will be
fixed in b75. If not, I will come back to your patch.
Tim
I think there's still an issue here that makes this patch worth
pushing. The 6888888 fix didn't cause the bug, but merely made it
visible to a lot more people. So 6889255 will only hide it again.
The build uses JAVA_TOOLS_DIR for javac, javah and javadoc:
# If no explicit tools, use boot tools (add VM flags in this case)
JAVAC_CMD = $(JAVA_TOOLS_DIR)/javac $(JAVAC_JVM_FLAGS) \
$(JAVACFLAGS)
JAVAH_CMD = $(JAVA_TOOLS_DIR)/javah \
$(JAVAHFLAGS)
JAVADOC_CMD = $(JAVA_TOOLS_DIR)/javadoc $(JAVA_TOOLS_FLAGS:%=-J%)
but only when LANGTOOLS_DIST is not defined. Normally, langtools is
built first so LANGTOOLS_DIST is defined. What your fix for 6888888
did was cause the setting for JAVAH to get used in all cases and, as
JAVA_TOOLS_DIR defaults to ALT_JDK_IMPORT_PATH rather than BOOTDIR
this fails in many cases as the user sets ALT_BOOTDIR not
ALT_JDK_IMPORT_PATH.
In jdk_generic_profile, it seems ALT_JDK_IMPORT_PATH is set to
${jdk_instances}/${importjdk} if it exists. On Solaris,
${jdk_instances} is set to /usr/jdk/instances which probably explains
why Sun engineers building on Solaris may not see this bug either.
${jdk_instances} is set to /opt/java on GNU/Linux, whereas distros
tend to have standardised on /usr/lib/jvm. It's simply "C:" on
Windows. As mentioned on IRC, release engineering are setting
ALT_JDK_IMPORT_PATH explicitly so also would never see this bug.
As we presumably want the tools from the bootstrap JDK, and not from
'the import JDK to be used to get hotspot VM if not built', I suggest
we switch to BOOTDIR by default by applying this patch.
This probably explains problems I've reported informally to Kelly a
while back, but never got around to formally investigating. FWIW,
this fix does not interfere with the fix for 6889255 coming up, and this
fix seems like a good idea to me.
-- Jon
