ARM Fixes

Few changes for ARM ThunderX2 and ARM-clang:

• Do not assume char is signed. Use int8_t for the time-bin .
• Improve the detection of the ARMv8.1a cycle counter assembly instruction.
• Fix cycle.h for ARMv8.1a. Will push upstream to FFTW.
• Add a synchronization point in the memswap() for clang on ARM.
• Do not let the autotools choose the flags for clang and GCC on ARM 8.1a. Use the native instructions.
• Upgrade to some of the autotools macros.

added 1 commit

• c7db28b4 - Fix present but cannot be compiled test for immintrin.h and altivec.h

Compare with previous version

what is that second test for? isn't what we have sufficient for the vector instructions?

Also cycle.h comes from fftw. we may want to push this upstream if that fix is confirmed.

I believe this is what the original fix was supposed to be (see the discussion on https://github.com/SWIFTSIM/swiftsim/issues/5) but for some reason that never got in?

Are you configuring with --enable-armv8-cntvct-el0 or --enable-armv8-pmccntr-el0 ?

added 1 commit

• 12d623a2 - Add the --with-arm-fftw flag to enable use of armpl_lp64 for fft. Should…

Compare with previous version

We're using --enable-armv8-pmccntr-el0. We can build with arm fft library now too.

@pdraper I somehow messed up the configure.ac here, if i do --with-arm-fftw=/path/to/lib then it randomly includes -Ltest/lib into the CFLAGS isntead of -L/path/to/lib and I can't work out why. It works fine if I just do --with-arm-fftw. It should jsut be the same as the with-fftw options, but I can't see where I went wrong :/. Any chance you can have a quick look?

if you are configuring with this flag then I don't understand why you need the change to the cycle.h file.

added 1 commit

• 4fb41190 - Fix up missing _arm parts of path

Compare with previous version

Just some missing _arm parts I think. Try that.

added 1 commit

• 98f8822a - Give a workaround for the issue seen with armclang and clang in memswap.h

Compare with previous version

changed title from Time-step limiter types to ARM Fixes

added 1 commit

• 977e6a88 - Correct use of vec_set1() in dimension.h

Compare with previous version

added 1 commit

• bc3f789c - Better fix to cycle.h for the ARM 8.1a architecture.

Compare with previous version

added 1 commit

• 73295cb2 - Do not look for the ARM FFTW library by default.

Compare with previous version

FYI I fixed cycle.h and the configuration script to still work on non-ARM systems.

added 1 commit

• 94a4bca8 - New upstream release of the ax_ext.m4 macros to detect vector instructions.

Compare with previous version

added 1 commit

• b4091b98 - Typo in the configure file for altivec instrinsics.

Compare with previous version

added 1 commit

• 0bb93ced - Type casting of the tic/toc before output to the task debugging files.

Compare with previous version

added 1 commit

• f99df277 - Fix arch for ThunderX2 to just use native

Compare with previous version