[0052] Finalize Experimental DXIL Op proposal

[V-FEXrt]

Presents the final proposed solution for experimental DXIL Ops which is to accept the 16bit FeatureID solution but to restrict the feature IDs to 0x0000 and 0x8000

[llvm-beanz]

I've commented on some spelling errors, but I think the core of this is good.

One concern I do have is that we've put a lot of words into this proposal that make it a lot less clear to come in with fresh eyes, read the proposal and understand exactly what was decided by it. This proposal isn't completely unique, but I think we've gotten to this position because there was extensive discussion about the alternatives.

We'll need to consider carefully in the future how we can balance keeping historical context but also being succinct in the technical documentation so that the record of what we're building is concise and digestible.

[Flakebi]

Doing a fresh read, I agree with @llvm-beanz, that a description of the actual accepted state is missing.
A final summary under “Proposed Solution” would be good to read and also to remove any ambiguities that come from the current description (which reads like “merge these two proposals in your head and add this third restricition”).
The descriptions are not long (the “Pros” would not belong into “Proposed Solutions”), the problem is that it’s scattered and contains two “diffs” that you need to apply in your head.

[V-FEXrt]

@Flakebi @llvm-beanz I simplified the "Accepted Solution" section to just state the facts of what we are doing as an independent proposal. Added a section just below on why that was the accepted proposal. Should be a lot easier now to just look at it to get an idea of the intended spec change.

[tex3d]

Looking good.

I pointed out a couple typos, some suggestions, and additional comments.

Nothing serious or blocking acceptance and merging.

[Flakebi]

Great, thanks for adding that. It’s much clearer now!

[llvm-beanz]

opcodes must be contiguous

I don't think this is technically true. The fact that DXC requires reserving opcodes is an oddity of DXC, but we could leave gaps, the only reason not to do that is that it has an undue cost to DXC.