`settag` should be atomic

[Adam-pi3]

When mc_tag_width = 4, two 4-bit tags share a single byte in VITT. A settag targeting a single chunk must read the byte, modify one nibble, and write the byte back. The spec explicitly states:
There are no atomicity requirements on the implementation for settag instruction.
That means, is it okay if hardware implementation with non-atomic RMW can override the other 4bit nibble when updating its own nibble in the same byte (lost-update issue)?

[Adam-pi3]

CC: Florian Mayer (@fmayer), deepak0414

[fmayer]

We agreed to keep like this and add an informative note to clarify that software should make sure that different harts do not manage the same 32 bytes (to type chunks). E.g. by making sure heap allocations and stacks are 32 byte apart (and aligned)

[deepak0414]

Concurrency issues in settag flow can come from

• Heap allocations coming from two different threads
• Stack allocations

Stack is private to thread and thus tag storage shouldn't see update from another thread for stack allocated to one.
In case of heap allocation, allocater can ensure atomicity via other methods.

If it's really required to have atomicity on settag, then I think ISA should enforce so.