anon (anon.delete@this.ymous.org) on September 15, 2021 12:15 am wrote:
[snip]
> Well, I don't think my concern was about permissions at all, but the discussion is nevertheless
> pertinent. I am not familiar with the PLB though, is there a reference ?

"A Comparison of Protection Lookaside Buffers and the PA-RISC Protection Architecture" (John Wilkes and Bart Sears, 1992, PDF) was my exposure to the concept. However, the principle is simply divorcing translation from permissions. This is somewhat obvious in that the granularity of permissions need not match the granularity of translation (and if one does not have aliases in a virtually tagged cache one can defer translation until time to access a memory-side cache or later with the obvious benefit of permissions being smaller (and potentially coarser-grained) than translations).

(There probably is potential for tag compression with temporal sharing of address references of permissions, translation, and data caching; a microarchitecture might choose not to fully isolate these aspects.)

(The intent of my four-word no-internal-text post was merely that permission checks do not require translation — not that permission checks should necessarily be deferred [and I do not need anyone's help to make an ass out of me].)

With respect to handling load-to-store forwarding, if there are no aliases/duplication in a virtually-tagged cache, then cache hit status could presumably be used to help manage store-to-load forwarding (in addition to not matching the page offset bits). Whether the complexity and area/power costs of managing aliases in a virtually-tagged cache are worthwhile for "normal" OSes that allow aliases is a different question than whether virtually-tagged caches are technically possible while supporting such OSes. Abnormal OSes are also theoretically possible even if the economics would not be favorable.