Trying to compile for ARM v7 on M1 runners

[jfalcou]

Hi,
I am the comaintainer of EVE, a C++ SIMD wrapper.
We wanted to migrate from our self hosted, cross compiling and qemu based ARM support to the new M1 runners.
It works perfectly for AARCH64 but I can't find a proper setup to compile for NEON only.

As EVE needs to be tested on all thing SIMD to ensure its quality, I was wondering if anyone had a way to do so.
Using the usual --target etc fails as it looks like the clang++ from homebrew just dont want to support it and looking at the log, looks like -arch arm64 is auto-added anyway.

Any hints appreciated.

[erik-bershel]

Hey @jfalcou!

I am not so familiar with this kind of edge case. I am not sure that M1 fully support old ARM v7 arch. But it should be available via cross-compilation at least, I suppose. So, I have at the moment just couple of silly ideas:

• to try to override CMake System Detection in toolchain file (dangerous area):

set(CMAKE_SYSTEM_PROCESSOR armv7-a)
set(CMAKE_SYSTEM_NAME Linux)

• to explicitly target on NEON with something like that (not sure about target):

set(CMAKE_CXX_FLAGS "-mfpu=neon -march=armv7-a -mfloat-abi=hard -target arm-linux-gnueabihf")

But two questions remain open - does M1 support cross-compilation for the old architecture and does LLVM contain libraries for such cross-compilation. In the second case, you will have to rebuild LLVM instead of the one installed from brew.