Add riscv64 as a supported manylinux architecture

[markdryan]

This will allow wheels with platform tags of the form manylinux_${GLIBCMAJOR}_${GLIBCMINOR}_riscv64 to be installed on riscv64 devices.

[brettcannon]

Should we worry about extensions? E.g. what will riscv64g show up as? And do re need to worry about it?

[markdryan]

Should we worry about extensions? E.g. what will riscv64g show up as? And do re need to worry about it?

I'm not sure whether this answers your question or not but,

• sysconfig.get_platform() returns "linux-riscv64" on a VisionFive2 running Ubuntu 23.10 with a 6.5 kernel. cat /proc/cpuinfo reports the following isa string for the VisionFive2 "rv64imafdc_zicntr_zicsr_zifencei_zihpm_zba_zbb".

• On QEMU instance running Ubuntu 23.04 with a 6.2 kernel, cat /proc/cpuinfo shows 'rv64imafdc' and sysconfig.get_platform() again returns "linux-riscv64".

[brettcannon]

I'm not sure whether this answers your question or not

Yes and no. It does in terms of it being an upstream problem. But it also makes me wonder if sysconfig.get_platform() is wrong?

[markdryan]

Yes and no. It does in terms of it being an upstream problem. But it also makes me wonder if sysconfig.get_platform() is wrong?

On Linux, the information for sysconfig.get_platform comes from the kernel via the uname syscall. The result returned on riscv64 seems to be consistent with other architectures. For example, on a AMD EPYC 7313 16-Core Processor running Ubuntu 22.04, sysconfig.get_platform() simply returns 'linux-x86_64'. The EPYC 7313 supports lots of CPU extensions, e.g., AVX, AVX2, but none of these are indicated via sysconfig.get_platform().

[brettcannon]

Sorry about the delay; PR got away from me.