[Android] Building mGBA with Clang

General Development
#1

Due to the removal of GCC in NDK r18, I was curious if I could get mGBA to compile with Clang and how it’d perform. I’m just posting my results in case anyone else was curious.

Here’s a patch that allows compilation with Clang: https://gist.github.com/GrieverV/a949c6f59577e3b0b6b2b955e7b17381

Patch notes:

  • Only armeabi-v7a’s Makefile was ported
  • Assumes you’re using a custom toolchain generated using make_standalone_toolchain.py --arch arm --api 22 --stl=libc++ --install-dir=$HOME/ndk-arm-22. You probably only need to pass -target and slightly change some paths to work with the default NDK toolchain.
  • ANDROID_TOOLCHAIN needs to be set to your toolchain path
  • -DHAVE_STRTOF_L -DHAVE_XLOCALE are now required
  • The DEBUG check always evaluated to true on my system so I changed it to better match similar check in Makefile.common which only enables DEBUG if explicitly set
  • futimes does not seem to be available but futimens is, added android ifdef to fix compilation
  • NDK r16(?) switched to unified headers which significantly simplifies includes
  • -mthumb -mtune=cortex-a72.cortex-a53 -mfpu=neon -flto -funroll-loops compiler flags were added
  • Due to using Android 22, -mthumb and -mfpu=neon, binaries produced using this patch likely require Android >=5.1 and a CPU that supports Thumb-2 and NEON.

Here’s a mGBA core produced using the above patch: https://mega.nz/#!VthgSawZ!8KoqEyiQTs1s60Iakyb4vmtGW9pdalwKqTgF2rT-wsM

libretro/mgba abf712f w/ upstream mgba-emu/mgba master (575adca) merged

I haven’t done much performance testing, but I believe performance was similar to the core produced by the buildbot. -mthumb appeared to give ~5% increase in fps and ~20% reduction in size. I haven’t tested with/without other flags.

1 Like
#2

Looked into futimes support and it appears futimens was added in API 19 and futimes was added in API 26. Locale support was also added around API 21, although I’m not sure to what extent.