This is based on a somewhat subjective review of the in-tree support, and where I thought further work was needed before I'd consider these "done".
Details
Diff Detail
Event Timeline
llvm/docs/RISCVUsage.rst | ||
---|---|---|
96 | This wasn't a deep dive. For Zbkb, I had observed we could use PACK variants for building constants or build_vectors of small vector types. Not clear how much we can/should push here. For Zbkx, we appear to have no pattern matching beyond intrinsics. |
llvm/docs/RISCVUsage.rst | ||
---|---|---|
96 | We could use PACK for assembling pieces of unaligned loads too. Zbkx pattern matching is missing and may be hard to do robustly. Maybe we if we did something like recognizeBSwapOrBitReverseIdiom where we collect the provenance of every bit. But that could still be fragile if it is proceeded by any extra bit operations. |
llvm/docs/RISCVUsage.rst | ||
---|---|---|
59 | +1, and also for the existing Zve32x/Zve32f/Zvl32b notes? |
Do folks mind if we do the link piece in a separate un-reviewed change? That's mostly a fight with rst formatting change, and I'd like to separate it from this.
Hm, I don't think this is happening. I tried playing with links, and it breaks rendering of the table. I have absolutely no idea why. If anyone else wants to play with this, you can see the result and the tweaked source here: https://github.com/preames/public-notes/blob/master/RISCVUsage.rst
Edit: After much banging of head into bring wall, I think I got this working in c6d304af.
Can we link this to the note?