Missing target feature ABI checks for Sparc64

[RalfJung]

The Target::abi_required_features and Target::check_consistency ABI-checking logic currently has no handling for Arch::Sparc64.

For context: on other targets this logic ensures things like

• valid values for llvm_abiname/llvm_floatabi
• consistent values for cfg_abi relative to llvm_abiname
• preventing target features from being enabled/disabled in a way that would alter the ABI

This is needed to prevent issues like #116344.

We have two tier 2 target for that architecture:

• sparcv9-sun-solaris
• sparc64-unknown-linux-gnu

And a few tier 3 targets:

• sparc64-unknown-helenos
• sparc64-unknown-netbsd
• sparc64-unknown-openbsd

To resolve this issue, we need to figure out whether there are any ABI differences between those targets, how LLVM determines which ABI to use, an how we can ensure that only code with a consistent ABI is ever linked together. The existing logic for abi_required_features and check_consistency is probably instructive to see how this is done for other targets.

Looking at the LLVM source code, the only "obviously" ABI-affecting target feature I was able to identify is the usual soft-float. So the "obvious" thing to do would be to mark that target feature as "forbidden" in abi_required_features, and maybe eventually add a soft-float target if anyone needs that. But maybe there is more?

Cc @psumbera @kulikjak @workingjubilee

[workingjubilee]

cc @jonathanpallant (I am aware that you maintain a 32-bit, not 64-bit SPARC target, but it seems likely your knowledge can inform this discussion. )

[workingjubilee]

cc @koachan (SPARC64 LLVM contributor, perhaps you know the answer to what target features can/should affect the ABI from the LLVM perspective? not all of our features are just a copy-paste of LLVM's but... a lot are, and we have a mapping, so!)

[taiki-e]

On 32-bit sparc, v8plus also affects the ABI, but on sparc64, AFAIK only soft-float does (#131799 (comment)).

[RalfJung]

On 32-bit sparc, v8plus also affects the ABI,

How does it affect the ABI?
I was not able to find a trace of this in the usual file in LLVM.

[taiki-e]

+v8plus changes it to use SPARC-V8+ ABI instead of the default SPARC-V8 ABI. At the time I last checked, this feature had not yet been fully implemented in LLVM (see description and FIXME in #132472), but this ABI makes a register reserved by the system in V8 ABI usable for users and extend the size of some registers (ABI doc).

[RalfJung]

Thanks. I hope this means it's just unimplemented and not that there's some other magic place where the ABI gets affected (as then who knows what other targets use that other magic place to adjust their ABI).

Should our targets use V8 or V8+?

[jonathanpallant]

My bare-metal 32-bit SPARC target is for the Gaisler LEON3, and that is a SPARC V8 processor.