cranelift meeting notes for 6-6-22 (#4235)

This commit is contained in:
Sam Parker
2022-06-07 16:56:48 +01:00
committed by GitHub
parent 20f510671d
commit 511c53703a

View File

@@ -17,4 +17,69 @@
### Attendees ### Attendees
- abrown
- akirilov
- avanhattum
- cfallin
- fitzgen
- sparkerhaynes
### Notes ### Notes
- ISLE porting status and plans?
- abrown: Quarterly goals at Intel, port a number of instructions over to ISLE.
- Want to continue doing so.
- Issue is that all the hard ones are left.
- We need more people to look at it.
- akirilov: Quarterly basis too.
- Our approach was different, we've gone for the harder ones first.
- Trying to improve test coverage.
- abrown: How are you guys tracking coverage?
- akirilov: It's just manual, the bar is currently quite low so easy to see where they're missing.
- Interpreter has failed to handle some cases, something to do with splat.
- Some run tests don't check the return values.
- Sometimes disabling the interpreter tests.
- cfallin: Sounds like the right approach.
- abrown: What's the state of SIMD interpreter support? afonso was working on it.
- cfallin: Don't know what happened to afonso, work has stalled on it.
- akirilov: Regalloc2 limitation of only 2 reg classes
- SVE will need a predicate file, same for AVX-512.
- cfallin: Bit packing issue, could remove the 'mod' operands and reuse that bit.
- akirilov: Will four classes be enough..? It is probably fine for AArch64.
- cfallin: Will keep this on my back burner.
# standups
- fitzgen: No updates.
- avanhattum: Better semantics for verification.
- SAIL for x64 and arm.
- Figuring out what work is needed to use modern SAIL for x64.
- Working on shim code to avoid annoting machine instructions.
- akirilov: cranelift CFI patch updates and Fiber changes for MacOS.
- bjorn3 gave feedback and noticed codegen wasn't right.
- some branches were elided and the branch target instructions ended up as the last instruction, not first.
- pointer authentication is still disabled in qemu.
- cfallin: Implemented alias analysis for load elimination, with okay speedups.
- could have benefitted some option optimizations, such as GVN
- So developing a unified framework for rewrite rules in the mid-end.
- egraphs seem the right approach.
- need an adapter to convert the egraph to LowerCtx: clif -> egraph -> vcode
- Then need to figure out isle rewrite rules.
- Framework will subsume all the existing mid-end optimisations.
- Also worked on a fix regalloc2 issue, reducing stack size usage.
- alexa: static rewrite rules or equality saturation?
- chris: equality.
- abrown: shared memory in wasmtime.
- impacts cranelift through `memory_size` instruction.
- How are we gonna test this? Looking for ideas and feedback.
- Ported some atomic operations for x64, not sure CAS is right.
- sparker: Dynamic vector RFC is up in code form.
- It's a bit rough and would greatly appreciate some feedback.
- cfallin: does qemu support sve?
- sparker: yes.