Update features tables to LLVM 21.1.8

[giordano]

CC @gbaraldi @topolarity (none of you is watching the repo...)

[giordano]

Don't quite understand what I'm supposed to do with the test failures. I frankly don't quite understand how to read the output, it's very confusing, there are lot of messages indicating "failures" and it's hard to tell what's relevant and what's not. I'm going to assume the relevant part is

 --- HW feature detection coverage ---
  FAIL: 'sve-sha3' is implied by a detectable HW bit but is not itself detectable (or baseline).  FAIL: 
  FAIL: 'sve-sm4' is implied by a detectable HW bit but is not itself detectable (or baseline).  FAIL: 
  FAIL: HW feature 'chk' is unhandled
  FAIL: HW feature 'sve-sha3' is unhandled
  FAIL: HW feature 'sve-sm4' is unhandled
  FAIL: HW feature 'use-fixed-over-scalable-if-equal-cost' is unhandled
  FAIL:     All HW features must be categorized: 
      - baseline (always present)
      - detectable (has a runtime probe implemented)
      - undetectable (no runtime probe, unsafe to enable)
      - featureset (only groups other features, no probe necessary)

  HW feature detection: FAILED

FAILED: 3 test(s) failed.

No idea what that "3 test(s) failed" refers to, it's impossible to understand what test fail, and what are those 3 (I see a lot more than 3 "FAIL" messages)

[topolarity]

CC @gbaraldi @topolarity (none of you is watching the repo...)

I only ended up here to fix some conspicuous bugs 🙂

Don't quite understand what I'm supposed to do with the test failures.

All of the failures are relevant.

If the tables are updated such that new features are added, the host detection logic needs to be updated to detect them so that we can safely enable them ("All HW features must be categorized"). That is your situation for chk, sve-sha3, and sve-sm4. On the other hand, use-fixed-over-scalable-if-equal-cost is being incorrectly categorized as a HW bit. If you get the detection resolved for the other features, I can help with that one.

[giordano]

All of the failures are relevant.

So messages like

--- CPU lookup test ---
  generic: 8 features
  haswell: NOT FOUND
  skylake: NOT FOUND
  znver4: NOT FOUND
  znver3: NOT FOUND
  nonexistent: NOT FOUND

are relevant? Would look like something bad at a quick glance, but I know this is running on aarch64 and not finding haswell/skylake/znver4/znver3 doesn't seem surprising, but then this is confusing ("what am I supposed to do with this?"), which is my point.

If the tables are updated such that new features are added, the host detection logic needs to be updated to detect them so that we can safely enable them ("All HW features must be categorized"). That is your situation for chk, sve-sha3, and sve-sm4.

Any hint of where the detection logic happens?

[giordano]

Ok, I asked for help to a friendly assistant, and at least tests are passing on this platform. It claimed that

• sve-sha3 and sve-sm4 are equivalent to sve2-sha3 and sve2-sm4, respectively, which seems plausible based on comments in LLVM source code
• chk and use-fixed-over-scalable-if-equal-cost both don't have any runtime probe on any platform, so it classified them as such. Not sure about this one, you said use-fixed-over-scalable-if-equal-cost may be wrong?

I also got it to add some more information to the README about what to do when upgrading LLVM version, with references.

I kept the changes in separate commits, to make review easier. Feel free to fix anything wrong!

[gbaraldi]

The sky lake etc messages are kind of sanity checks for doing things locally (and just to know what cpu CI might have) but thanks for reminding me that I need if defs around them

[topolarity]

The use-fixed-over-scalable-if-equal-cost being marked as a HW feature is indeed an LLVM bug, already fixed in trunk by llvm/llvm-project#152156. I blacklisted it for now, which should have the same effect.

It sounds like chk is effectively part of the baseline feature set on all AArch64 platforms:

Since the CHKFEAT instruction is in the Hint instruction encoding space, it is effectively implemented on all PEs that support AArch64

https://developer.arm.com/documentation/ddi0487/mb/-Part-D-The-AArch64-System-Level-Architecture/-Chapter-D1-The-AArch64-System-Level-Programmers--Model/-D1-11-Check-Feature

Pushed 2 commits to fix each. I think the host detection is OK now

[topolarity]

Took the liberty of rebasing on top of #25.

I think this is ready to merge, assuming CI is happy.

[topolarity]

Thanks for kicking this process off @giordano !

[giordano]

Thanks for pushing it to the finish line 😃