Always specify llvm_abiname for RISC-V targets

[beetrees]

For RISC-V targets, when llvm_abiname is not specified LLVM will infer the ABI from the target features, causing #116344 to occur. This PR adds the correct llvm_abiname to all RISC-V targets where it is missing (which are all soft-float targets), and adds a test to prevent future RISC-V targets from accidentally omitting llvm_abiname. The only affect of this PR is that -Ctarget-feature=+f (or similar) will no longer affect the ABI on the modified targets.

r? @RalfJung

[rustbot]

These commits modify compiler targets.
(See the Target Tier Policy.)

[RalfJung]

I can't speak to the consequences of this PR, so I can't review it, sorry. All I am doing is going around and bothering everyone to get their ABIs in check, I don't actually know the technical details of how exactly to do that on each architecture. ;)

r? compiler

[RalfJung]

The only affect of this PR is that -Ctarget-feature=+f (or similar) will no longer affect the ABI on the modified targets.

That sounds great!
What will -Ctarget-feature=-f do when the abiname says "use floats"?

[workingjubilee]

What will -Ctarget-feature=-f do when the abiname says "use floats"?

I think the answer is actually, surprisingly, that LLVM shrieks "DON'T!" at us. Obviously we have to actually reach codegen, but LLVM actually checks this sort of dependency for the RISCV backend.

[RalfJung]

I wish that would be true. However, according to this, the answer is "Disabling target features required by the requested ABI will cause LLVM to ignore the ABI", which sounds less nice.

[RalfJung]

Also we probably want to show a nicer error than have LLVM do error reporting. But it's always good to have a safety net :)

[workingjubilee]

Ah, yeah, I checked and the code that I thought might be there isn't. Sadness!

[RalfJung]

Would it be worth asking LLVM to add such checks? For the arm backend at least that might actually be happening.

[RalfJung]

See <llvm/llvm-project#110383> for the arm issue.

[workingjubilee]

I think we can make that happen, yeah. RISCV does some forms of feature dependency-checking already.