Mark c0-authorized loads and stores as extensible

[nwf]

Say explicitly that c0-authorized load/store instructions, which are guaranteed to trap regardless of other operands' values, are available for use by future extensions.

[jrtc27]

Why? They’re not for the base ISA.

[nwf]

They stand a chance of working in the base ISA (ETA: or in Zcherihybrid), if there's memory near 0. This is specifically for instructions in cap mode.

[arichardson]

If you plan to repurposed this "dead" opcode, I wonder if it would be better to have them raise an illegal instruction? Although that would require special casing in the decoder so maybe the current behaviour makes more sense?

[tariqkurd-repo]

@nwf is pointing out that if the auth cap is c0 you definitely get a tag violation, allowing future extensions to redefine that case, as oppose to it having to be a tag violation for all time.

Do we need to cal this out explicitly? I suppose we do to prevent it always being a tag violation.

[nwf]

@arichardson Yeah, I agree that in some sense "illegal instruction" is "better" than "CHERI tag violation" if the intent is to allow the instruction to be repurposed (and yes, indeed), but I don't think we need to mandate that complexity in the decoder. It's perfectly fine, IMHO, for a Zcheripurecap core to really try operating on the NULL cap here while also allowing extensions to operate on something else (by way of concrete examples, perhaps, DDC like we did back in CHERI-RISC-V or the Zstid *tidc registers).

@tariqkurd-repo I do not know how RISC-V in general feels about extensions replacing instructions without permission of the specifications they extend, but that seems kind of rude. It seemed more polite, if nothing else, to explicitly grant permission to do so. :)

[tariqkurd-repo]

I'll note that there is more redundancy in the encodings as the load/store addresses must be 16-byte aligned and so the 4 LSBs (1/3) of the immediate are wasted, as well as c0 being useless.

It would be better to make all of those cases reserved.

I personally think these things will come out of ARC review when we get to that stage, and they allocate encodings.

[davidchisnall]

I'll note that there is more redundancy in the encodings as the load/store addresses must be 16-byte aligned and so the 4 LSBs (1/3) of the immediate are wasted, as well as c0 being useless.

This is not wasted space. The alignment of the address + immediate needs to be 16-byte aligned, but there's no guarantee that the address is already 16-byte aligned.

In particular, it is very common in managed languages to have tagged pointers where the low 3 bits store type information. This can then be reset to zero in the load / store instructions with a -1 to -7 displacement added to the field offset.

[nwf]

Just to make sure it isn't missed... In addition to the clever trick David's pointing out, this marking as extensible is intended to apply to more than just high-alignment-requiring LC/SC instructions (including LB/SB, which impose no alignment requirements).

[tariqkurd-repo]

thanks @davidchisnall I didn't know those bits would be used that way, so if they don't offer those bits when we finally get allocated encodings, we should point to that use case

@nwf I guess you're saying that load/stores via c0 can be replaced by anything you like

[nwf]

@tariqkurd-repo Yes, though I imagine they'll be replaced with load/stores still, since all the decoder logic is right there, but in principle...