[Toybox] toys/android/log.c with android NDK?

[enh]

On Mon, Mar 12, 2018 at 11:06 AM, Rob Landley <rob at landley.net> wrote:
> On 03/10/2018 11:06 PM, enh wrote:
>>> $ sudo build/tools/make_standalone_toolchain.py --arch x86_64 --api 26 \
>>>   --stl=libc++ --install-dir=/opt/android/x86-64
>>> $ CROSS_COMPILE=/opt/android/x86-64/bin/llvm- CFLAGS=--static make defconfig
>>> It has no llvm-cc... ok then:
>>> $ sudo ln -s clang /opt/android/x86-64/bin/llvm-cc
>>> $ CROSS_COMPILE=/opt/android/x86-64/bin/llvm- CFLAGS=--static make defconfig
>>> $ CROSS_COMPILE=/opt/android/x86-64/bin/llvm- CFLAGS=--static make
>>> many warnings ignored...
>>> generated/obj/log.o:log.c:function log_main: error: undefined reference to
>>> '__android_log_write'
>>> $ CROSS_COMPILE=... make menuconfig # switch off log in android menu.
>>
>> (or add -llog.)
>
> Hmmm, should already be there in scripts/make.sh...
>
>   # We trust --as-needed to remove each library if we don't use any symbols
>   # out of it, this loop is because the compiler has no way to ignore a library
>   # that doesn't exist, so we have to detect and skip nonexistent libraries
>   # for it.
>
>   > generated/optlibs.dat
>   for i in util crypt m resolv selinux smack attr rt crypto z log
>   do
>
> Ah, here's the problem:
>
> $ /opt/android/x86-64/bin/llvm-cc -llog hello.c
> $ ./a.out
> bash: ./a.out: No such file or directory
> $ /opt/android/x86-64/bin/llvm-cc -llog hello.c --static
> /opt/android/x86-64/bin/../lib/gcc/aarch64-linux-android/4.9.x/../../../../aarch64-linux-android/bin/ld:
> cannot find -llog
> clang60: error: linker command failed with exit code 1 (use -v to see invocation)
>
> The library only exists for dynamic, not static. And I can't run a dynamic
> bionic binary on an ubuntu host.

not _strictly_ true: you can install /system/bin/linker64 and all the
libraries you need in /system/lib64.

> Of course...
>
> $ /opt/android/x86-64/bin/llvm-cc --static hello.c
> $ ./a.out
> Illegal instruction (core dumped)
>
> Still an issue with api 28.

works for me.

are you running on hardware so old that it doesn't meet the minimum
ABI requirements for Android x86-64? from
https://developer.android.com/ndk/guides/abis.html that's MMX, SSE2,
SSE3, and SSSE3, SSE4.1, SSE4.2, and POPCNT.

if you're seeing SIGILL on an instruction not covered by those, that
would be a bug. but Android never shipped on CPUs lower than the above
bar.

>>> $ CROSS_COMPILE=... make
>>> more warnings, more warnings...
>>> scripts/make.sh: line 28: /opt/android/x86-64/bin/llvm-strip: No such file or
>>> directory
>>> strip failed, using unstripped
>
> You know that one's trivial enough I might just add a "strip" to toybox. (It's
> basically objcopy with some excludes.)
>
> (The hiccup is that the japanese developers implementing the sh4 toolchain had a
> translated copy of the ELF spec and the codepage switch turned "_" into "." so
> they dutifully used . as the symbol prefix, and if you use the wrong strip there
> it mangles your executables. So one strip working on all targets requires
> checking to see what if any prefix is currently in use, and that's why I use the
> prefixed strip.)
>
>>> And that file is doing #include <iconv.h> so I don't see _why_ it's not getting
>>> them?
>>
>> because <iconv.h> isn't there until API 28, but a side-effect of the
>> single sysroot for all API levels is that the *file* is there, but the
>> contents are #ifdef'ed out since you asked for API level 26.
>
> Cut and paste from the example in the web page. :)
>
> I redid it with api 28, and switching off android/log I got the rest of it to
> build. But the result still immediately segfaults, as does hello world.
>
> (I tried building with the included gcc toolchain too: same thing.)
>
>> we'll be able to provide better errors (along the lines of "iconv_open
>> is only available in API level 28 and later") when we remove GCC later
>> this year and only have to support clang, but r17 still includes GCC
>> so it's just #ifdef'ed out.
>
> The ability to build an x86-64 hello world binary statically linked against
> bionic which runs on an ubuntu host would be nice. Is there a design reason
> that's not supported?
>
> Rob