RFR: 8370659: [mobile] enable static-libs-image for Android

[Johan Vos]

On Sun, 26 Oct 2025 16:10:09 GMT, Johan Vos <jvos at openjdk.org> wrote:

> This PR adds the changes required to compile libjvm.a and the classlibs using the latest upstream code.
> 
> Some notes:
> 
> To configure/make, the following configuration is recommended:
> 
> bash ./configure \
> --disable-warnings-as-errors \
> --openjdk-target=aarch64-linux-android \
> --with-boot-jdk=/opt/jdk-25 \
> --with-toolchain-path=/opt/android-ndk-r29/toolchains/llvm/prebuilt/linux-x86_64/bin \
> --with-sysroot=/opt/android-ndk-r29/toolchains/llvm/prebuilt/linux-x86_64/sysroot \
> --with-toolchain-type=clang
> 
> 
> After configuring the build, run
> `make static-libs-image`
> 
> Note that this requires Android NDK 29, which can be downloaded from https://developer.android.com/ndk/downloads 
> I only tested this using Linux. Mac/Windows builds may or may not work.
> 
> In `make/autoconf/flags-cflags.m4`, the minimum android version is set to 32 (`-target aarch64-linux-android32`). The reason for this is that this version comes with a version of bionic that supports thr_current. This is not strictly needed, but lower versions would require more bionic-specific patches.
> 
> I use `#ifdef  __BIONIC__ `and similar in most places where we do "Android" specific checks. The reason for this is that those checks mainly refer to support that is (or isn't) available in Bionic. As such, it is better to compare e.g. bionic with musl and libc unless we are specifically using Android API's
> 
> I can provide a sample with a Makefile that validates the build on a real phone, but I think that is not in scope of this PR.

> I believe you can use the touch command to keep a Pull Request open nowadays, it's quite a nifty tool to refresh the Pull Request without those keep alive comments :)
> 
> I'm not a Reviewer, so can't really help with getting this integrated, but something general I've noted is the #ifdef code for **BIONIC** seems to not have a replacement for Android, it just doesn't include the code it guards on Android, but also doesn't implement it for Android either. So there's functionality in Java that is unimplemented and will simply not work on Android. 

That is true. There are some locations where #ifdef BIONIC is used to guard unneeded code on Android, but there are other locations where it is used to not execute some specific functionality that is not (yet?) working on Android. Some of this code doesn't make sense on Android, where other code probably should be implemented one day when we want to pass the TCK. I see this as part of a gradual approach, where we first want to make sure HelloWorld runs, and then add missing functionality.

> Also, the build system seems to have "android" properly set in some areas (Like OPENJDK_TARGET_OS) while in others it re-uses "linux", which makes it a bit harder to see what is for Android and what is for Linux. Are both intentional? Otherwise, this seems fine. Just a couple other not-too-serious queries.

Right, that is something I'm not too happy with myself, but I don't see a better solution. In the typical OS systems, we use e.g. macosx, linux, windows, aix as OPENJDK_TARGET_OS and then either unix or windows as OPENJDK_TARGET_OS_TYPE. I believe it's clear that we have to use unix as the OPENJDK_TARGE_OS_TYPE, but I've set the TARGET_OS to Android and not to linux, as there are important differences. However, in some cases we want to leverage a whole part of the linux code, and in that case we claim to be "linux". Ideally, there would be a 3-level system (e.g. Android -> Linux -> Unix) but that would introduce much more complexity.

I'm open to suggestions of course.

-------------

PR Comment: https://git.openjdk.org/mobile/pull/40#issuecomment-3686025022