Closed Bug 1710882 Opened 7 months ago Closed 2 months ago

SIMD shuffle optimizations for reverse bytes ops

Categories

(Core :: Javascript: WebAssembly, enhancement, P3)

ARM64
All
enhancement

Tracking

()

RESOLVED FIXED
95 Branch
Tracking Status
firefox95 --- fixed

People

(Reporter: yury, Assigned: yury)

References

(Blocks 2 open bugs)

Details

Attachments

(1 file)

The ARM64 has VREV16/VREV32/VREV64 permute operations that are used in the wild. It will be nice to add them to the AnalyzeShuffle and generate single instruction for ARM64 (and somewhat efficient code for x86)

This seems like it doesn't need to make the train for FF90, so lowering priority.

Severity: S2 → N/A
Priority: P1 → P3
Blocks: 1695599
No longer depends on: 1695599
Assignee: nobody → ydelendik
Status: NEW → ASSIGNED

There was a win at reverseInt16x8 about 8% on my local benchmark on a x64 machine.

On ARM64 platform, the local benchmark test 1-2% win for all added reverse ops.

Pushed by ydelendik@mozilla.com:
https://hg.mozilla.org/integration/autoland/rev/d453591e1259
Optimizations for SIMD reverse bytes ops. r=lth
Status: ASSIGNED → RESOLVED
Closed: 2 months ago
Resolution: --- → FIXED
Target Milestone: --- → 95 Branch
You need to log in before you can comment on or make changes to this bug.