On 2016-01-26 20:50, Bob Vandette wrote:
Please review the changes for two improvements for the Mobile Dev Forest.
With these changes, the mobile/dev forest can successfully build x86 and ARM
Android
JDK 9 binaries. The ARM binaries use the Zero ARM interpreter while the x86
implementation uses standard x86 Hotspot VMs.
The webrev also includes 1 fix for iOS BSD builds in
hotspot/make/bsd/makefiles/vm.make.
JDK-8148304 Mobile: iOS builds fail on assembler files
https://bugs.openjdk.java.net/browse/JDK-8148304
JDK-8148302 Mobile: Initial bring of Android x86 and ARM
https://bugs.openjdk.java.net/browse/JDK-8148302
Here’s the webrev:
http://cr.openjdk.java.net/~bobv/8148302/webrev
Hi Bob,
Some comments:
* In jni_util_md.c, you have accidentally lost a trailing ).
* In spec.gmk.in, please use ":=" instead of "=". The former is
assignment, the latter is used for macro expansion. (As a rule of thumb,
:= should always be used unless you have explicit reasons.)
* In CompileJavaModules.gmk, why do you remove the exclude for
BUILD_HEADLESS_ONLY? This change affects all platforms, not only android.
* In lib-x11.m4, I *really* don't like adding X11 paths when we should
not have X11! (And it's even worse since you do not check for Android
but do so for all platforms.) What files are needed, and for what
library? Can you not fix this dependency by removing the #include
statements in the source code instead? Otherwise, I'd say that android
*is* indeed dependent on X11, and you need to express this properly.
* In lib-ffi.m4 and Copy-java.base.gmk: I think this a bit problematic.
I believe what you're trying to achieve is similar to what we do for
freetype, where we "bundle" the library with the resulting product. I'd
rather see this libffi bundling based on that behavior. More concretely:
- The name of the variable should be like LIBFFI_BUNDLE_LIB_PATH, to
clearly show that this is used only when bundling the product, not when
compiling it.
- Whether to bundle or not should be determined in lib-ffi.m4. The
default value should be "bundle if android+zero, otherwise not". If you
want to, you can add a --enable-libffi-bundling to override this. I
won't require it, but I generally find such things useful since you can
test that bundling works even without building for a specific target.
The code in Copy-java.base.gmk should then only test if this variable is
present, compare with freetype in Copy-java.desktop. Finally, the
configure code must check for the case that libffi bundling is enabled
but no explicit libffi path is set (e.g. it is autodetected). It's okay
to abort with an error, but it must be detected.
And finally, a note of appreciation. :-) I'm really happy to see the
ugly solution with the hard-coded linker scripts go away in flags.m4!
/Magnus
