bjorn3 commented on Issue #1424:
https://github.com/bytecodealliance/wasmtime/issues/1423#issuecomment-605295620
Could you add an x86_pextr binemit test
Is the last commit ok, or do you want to test more cases?
github-actions[bot] commented on Issue #1424:
Subscribe to Label Action
This issue or pull request has been labeled: "cranelift", "cranelift:meta"
<details> <summary>Users Subscribed to "cranelift"</summary>
- @bnjbvr
</details>
To subscribe or unsubscribe from this label, edit the <code>.github/subscribe-to-label.json</code> configuration file.
bjorn3 commented on Issue #1424:
With this I can finally store SIMD vectors in SSA values, rather than memory. This improves performance for tokei from ~50% overhead compared to cg_llvm to ~30% overhead.
abrown commented on Issue #1424:
I don't know if this will give you more flexibility, but note that in #1377 I have added support for declaring constants in the function preamble instead of only through
vconst
. This also should make it easier to use constants of sizes other than 128-bit. Take a look at
64c8cbf and let me know if you can think of any changes to that that would help cg_clif.
abrown edited a comment on Issue #1424:
I don't know if this will give you more flexibility, but note that in #1377 I have added support for declaring constants in the function preamble instead of only through
vconst
. This also should make it easier to use constants of sizes other than 128-bit. Take a look at 64c8cbf and let me know if you can think of any changes to that that would help cg_clif.
bjorn3 commented on Issue #1424:
I don't think any of that PR will help me. At least for now. Creating SIMD values is currently very slow anyway as at MIR level all lanes get assigned separately, meaning that I have to use
vconst 0
+insertlane
. I will obviously need to optimize this :)
Last updated: Oct 23 2024 at 20:03 UTC