RFR (XL): JDK-8204572 SetupJdkLibrary should setup SRC and -I flags automatically
Magnus Ihse Bursie
magnus.ihse.bursie at oracle.com
Thu Jun 7 20:22:48 UTC 2018
This change needs some background information.
I've been working on simplifying and streamlining the compilation of
native libraries of the JDK. Previously, I introduced the
SetupJdkLibrary function, and started to get a better control of
compiler flags. This patch continues on both paths.
The original intent of this change, which I naively thought was going to
be much simpler than it turned out :-) was to separate the -I flags
(location of header files) from other flags, and to generate these
automatically, wherever possible. Since we have a standard way of
locating native code, and most libraries just want to have an -I path to
their own source base and the generated Java header, we should be able
to provide this in SetupJdkLibrary.
This turned out to be closely related to SetupJdkLibrary being able to
discover the location of the SRC directories itself, using "convention
over configuration" and assuming that the library "libfoo" for
"java.module" would be located in java.module/*/native/libfoo.
While this sounds simple in theory, when actually trying to implement
this I of course ran into all the places where some special handling was
indeed needed. So even if like 90% of all libraries were simple to get
to build using automated discovery of source and header directories, the
10% that did not caused me much more headaches than I had anticipated.
On the other hand, now that I've sorted out all those places, the few
remaining odd solutions is clearly documented and not just something
that "just happens" due to strange configurations.
One file deserves mentioning specifically: Awt2dLibraries.gmk. The
java.desktop libraries are unfortunately quite entangled with each
other, and do not readily follow the patterns that are used elsewhere in
the code base. So it might just look like the file has just gone from
one state of messiness, to another, which would hardly be an
improvement. :-( I would still argue that the new messiness is better:
It is now much clearer in what ways the libraries diverge from our
standard assumption, and what course of action needs to be taken to
minimize these differences. (Which is something I believe should be done
-- these issues are not just cosmetic but is the root of most of the
issues we always see for these libraries, when upgrading compilers, etc.)
During this change, I noticed that not all native libraries include the
proper generated header file. This is a dangerous coding practice, since
a change in the Java part of the interface might not get picked up
properly in the native part. I've added the missing includes that I've
detected, and due to these changes, I'm also including the component
teams in what is really only a build change. As can be seen for
jdk.crypto.mscapi; there had indeed been changes that needed correcting.
Since this is (basically) a pure build change, my gold standard here has
been the build compare script. In essence, the build output prior to my
change and with this change are 100% identical. In truth, this is a bit
too strong a claim. A few changes has occurred, but none of them should
matter. Here's a breakdown of the compare.sh results:
* Windows-x64:
No differences at all.
* Solaris:
Two libraries are reported to differ: libsaproc.so and
libfontmanager.so, both with a disass diff on ~700 bytes. Analyzing
this, I found that the object files used to link these two libraries has
no disass differences. They have a slight binary difference and a
difference in size, due to the include paths being different (and this
is stored in the .o file, which makes it different). Somehow this
apparently triggers the linker to generate a slightly different code in
a few places, using a different register or so. (Weird...)
* MacOS:
Two libraries are reported to differ: libjava.dylib and
libmlib_image.dylib, both of them just reported as a binary diff (no
symbol, disass or fulldump differences). This is not really unsuspected,
but I analyzed it anyway.
I found that for libjava.dylib, a single .o file was different. This one
was actually picked up from closed sources, and are not really relevant
for OpenJDK. Anyway, the reason for the difference was the same as for
the Solaris libs; the include paths had changes, which caused a binary diff.
For libmlib_image.dylib, the link order had changed causing the noted
binary difference.
* Linux:
On linux, the compare script noted differences for libextnet, libjava,
libmlib_image, libprefs, libsaproc, libsplashscreen and libsunec.
The differences for libextnet, libprefs, libsplashscreen and libsunec
turned out to be caused by the added #include of the generated Java
headers. This caused binary differences (reasonably), and for some odd
reason also a symbol difference in java_awt_SplashScreen.o (clazz.10057
and mid.10058 were replaced by clazz.10015 and mid.10016). I can't claim
to understand this, but I'm assuming it's some kind of generated code.
libsaproc and libjava changes was caused by closed source changes, and
is therefore not relevant to OpenJDK.
For libmlib_image.dylib, the link order had changed causing the noted
binary difference, as on MacOS.
Bug: https://bugs.openjdk.java.net/browse/JDK-8204572
WebRev:
http://cr.openjdk.java.net/~ihse/JDK-8204572-autodetect-SRC-and-headers-dirs/webrev.01
/Magnus
More information about the build-dev
mailing list