Consider weakening return sentries to support compiler outlining

[rmn30]

As per CHERIoT-Platform/llvm-project#46 the current restrictions on return sentries cause difficulties for compiler outlining because it may use a non-standard link register. Currently cjal / cjalr always produce a return sentry in the link register which can only be used to return via cret (i.e cjalr with cd == $cnull and cs1 == $cra).

For reference here is the relevant table for cjalr from the ISA doc:

We could relax the ISA to permit using a non-standard link register by modifying cjal / cjalr where cd is not $cra such that they produce either 1) an unsealed capability or 2) a forward interrupt-inheriting sentry in the link register. This would permit a return using cjalr with cd == $cnull and cs1 != $cra. Of these 1) is preferable as it does not introduce the ability to create forward sentries, which is not currently possible without a capability with the appropriate sealing permission. It is already possible to create unsealed capabilities derived from PCC.

[davidchisnall]

If we made a non-cra link register provide an unsealed capability, then the following two sequences are equivalent:

  cjal {target}, ct0

0:
  auipcc ct0, %cheriot_compartment_hi(1f)
  cincoffset ct0, ct0, %cheriot_compartment_lo_i(0b)
  j  {target}
1:

In both cases, you return with cjr ct0, which is already permitted.

I believe this means that there is no security implication and the relaxation on cjal[r] simply allows a code-size reduction by folding three instructions into one.

[rmn30]

As David points out this doesn't really weaken the current model at all. It really just fixes an anomaly in the above table where row three is not currently that useful: it can be used for making a function call with a non-standard link register but the called function has to move that register into $cra in order to return to it, which probably defeats the point of using a non-standard link register. The important thing is that calls to interrupt enabling / disabling sentries must use $cra so that the called function can definitely restore the previous interrupt stance.

[vmurali]

Another thing to note is that, in a call to a library, if a caller compartment wishes to sanitize the arguments and force the library to only use the correct return entry points (i.e. not trust the library), it can always use $cra as the link register in the call.

[resistor]

I did a measurement with and without outlining disabled. Re-enabling MachineOutliner will save us approximately 4.4% code size (according to scripts/output_code_size.sh) on the CheriotRTOS test suite binary. As such I think it's definitely worthwhile to push ahead on getting this change in.