RVM23: consider inclusion of Zilsd/Zclsd

[Wren6991]

Per title. These extensions have a significant impact on code density (static + dynamic), and the 64-bit accesses can be cracked into pairs of 32-bit accesses on low-end implementations. I think the main reason for them not being included is that they were only recently ratified, so weren't on the table when the original draft was drawn up.

For example, comparing Zilsd + Zclsd to the currently mandatory Zcmt, and assuming a 32-bit bus implementation:

• Similar static code size benefit (embench)
• Zilsd + Zclsd are performance-neutral; Zcmt is typically performance-negative without a BTB to cache the lookups
• Zilsd + Zclsd have lower gate count than Zcmt (the main culprit being the jvt CSR and the muxing it implies)

They're also particularly useful on RV32E, which frequently does not benefit from Zcmp compression due to lack of s registers (and reduced stack alignment), but can make use of paired load/stores to spill and fill a and t registers over a call.

I'm assuming that any Zilsd implementation would also support Zclsd, since all of the Zclsd instructions are simple aliases of Zilsd instructions, and RVM23 already mandates some amount of compressed instruction support.

[aswaterman]

Given that RVM needs to have fairly broad applicability but is a bit less concerned with binary compatibility than RVA, I think these should be expansion options rather than mandatory.