Add toolchain conventions for loongarch32

[jiegec]

No description provided.

[jiegec]

I suggest making la32r as a ISA subset of la32, or la32 begin a ISA superset of la32r. They are distinguished by -march=loongarch32 and -march=loongarch32r. What's your opinion?

[chenhuacai]

I suggest making la32r as a ISA subset of la32, or la32 begin a ISA superset of la32r. They are distinguished by -march=loongarch32 and -march=loongarch32r. What's your opinion?

I agree, you can start coding now. Code can help us to confirm the correctness and get this PR merged finally.

[xen0n]

Yeah I would agree to not distinguishing LA32S and LA32R via ELF e_flags or something similar, at least not right now. Starting from simple -march configuration would be most simple.

Eventually we may want to mark object files the RISC-V way, via a dedicated ELF attribute section, to avoid accidental ISA/ABI mismatches. But a naive approach e.g. just recording the ISA features used and disallowing interlink if linking for a lower baseline, would not allow for easy implementation of best-practices such as "dynamic ISA probing via HWCAP/CPUCFG + dispatch to optimized code" -- without post-processing of object files, the resulting program would appear as if it required the extra ISA features to run. Hence, a proper scheme that allows for both:

• accurate description of baseline requirements, and
• easy implementation of dynamically-dispatched optimization,

would take more time to get designed and PoC made, and I'd prefer prioritizing the current LA32(R) work over that.