Add meeting notes for 2022-06-13 Cranelift meeting (#4263)

Feel free to add suggestions / amend wherever is needed.
This commit is contained in:
Benjamin Bouvier
2022-06-13 18:13:26 +02:00
committed by GitHub
parent 1df731d5c2
commit 3eb5ece625

View File

@@ -17,4 +17,60 @@
### Attendees
- abrown
- akirilov
- avanhatt
- bjorn3
- bnjbvr
- cfallin
- fitzgen
- jlbirch
- sparker-arm
### Notes
Agenda item:
- bnjbvr: who owns the cranelift-jit-demo repo? is it collective responsibility? meta
question: how do we decide that in general?
- cfallin: meta question, probably a bytecode alliance thingy, RFC to discuss.
For this particular repo, would github's CODEOWNERS work?
- bnjbvr: explains how CODEOWNERS works. Not a solution for merge right. Github
has different allowance access for individuals, so could give merge rights to
a specific individual.
- cfallin: either open an RFC, or discuss this at next wasmtime meeting
Updates:
- sparker-arm: aarch64 vector work, benchmarking, moving along well, no PR opened yet
- cfallin: worked with egraph, subsumes GVN now, LICM soon. Question is how
does the rewrite system look like? Now that something works, write up an RFC
and see what people think.
- avanhatt: lots of verification updates, half way reviewing egraph PR
- bnjbvr: work paused the last two weeks on incremental cache, back to it this week
- akirilov: PAC (pointer authentication) work in fiber has been merged, CI uses PAC now, RFC + patch to
be updated soon
- cfallin: that's on linux aarch64, how far are we from enabling for mac
m1?
- akirilov: codegen changes was easy, unwinding harder, not sure about
the complexity for mac m1.
- bjorn3: mac m1's ABI is slightly different from linux aarch64's one
- jlbirch: talked about security concerns with Alex re: sightglass /
benchmarking.
- cfallin: how will this work? bot? manual trigger?
- jlbirch: anyone with sufficient permissions can comment/open issue to run
a workflow run (repository dispatch event), private repo will run the
PRs, do the testing and send the results back to the PR/issue. This will
be running on private machines (linux x64, linux aarch64).
- sparker-arm: limit egraph optimizations, how does it work?
- cfallin: fuel mechanism to limit # (number of) rules of application, metric on
memory usage (# nodes, classes), hard cap this to some multiple of #
cranelift input nodes.
- sparker-arm: (*notetaker missed that question*)
- cfallin: no rewrites at all at the moment, just hash-const. Turn up knob to
get several optimization rounds.
- avanhatt: would we do inter-procedural analysis in the future? ie allow
rules to rewrite across function boundaries somehow?
- cfallin: prob not. Could blow up memory/time, so we'd need to explore.
We'd do that only if we inlined that callsite already (so not across
function boundaries)