On Mon, 11 Jul 2022, Martin Storsjö wrote:

Hi,

Thanks for your patch! While the patch probably is worthwhile to pursue, ffmpeg does work on Android 6 and newer (with the aarch64 assembly), so there's some gaps/misses in the reasoning in the patch description.

On Sun, 10 Jul 2022, Triang3l wrote:

To support PIC, all AArch64 assembly code in FFmpeg uses the `movrel` macro to obtain addresses of labels, such as lookup table constants, in a way that with CONFIG_PIC being 1, PC-relative addresses of labels are computed via the `adrp` instruction.

This approach, however, is suitable only for labels defined in the same object file. For `adrp` to work directly between object files, the linker has to perform a text relocation.

No, it doesn't. It's fully possible to build such libraries without text relocations currently.

My memory of the situation was that we're linking our shared libraries with -Wl,-Bsymbolic, which means that those symbol references are bound at link time, so the offset from adrp to the target symbol is fixed at link time, and no text relocation is needed.

However I did try to link such shared libraries, removing the -Wl,-Bsymbolic argument, and it still succeeds. So I'm a bit unsure what really makes it work for me when it isn't working for you.

Andreas Rheinhardt kindly reminded me that when linking libavcodec.so, we add -Wl,--version-script,libavcodec/libavcodec.ver, which makes e.g. ff_cos_32 a non-exported symbol, which means that it can't be interposed, and thus the relocation can be fixed at link time.

If I remove that argument, I can reproduce the linker errors, and adding -Wl,-Bsymbolic fixes it.

I wonder if we could mark these symbols as ELF hidden, so that they wouldn't need to be interposable at all, even when you link the static library into a separate shared library?

// Martin
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".

Reply via email to