1762413 - Translate emscripten's sequence for _mm_maddubs_epi16 into MozWHPMADDUBSW

Status: RESOLVED FIXED

(Core :: JavaScript: WebAssembly, enhancement, P2)

Description

[Yury Delendik (:yury)]

Here is the sequence: https://github.com/emscripten-core/emscripten/blob/9feae9e946a42e60e9d8b6cd446a7cb8f0563f8d/system/include/compat/tmmintrin.h#L82-L90

There are saturated add+four shifts+two muls+and+const, but it is doable to match it during MIR's foldsTo. There is a very low risk that compilers will replace this with different operations during optimization.

This will allow us to provide fast PMADDUBSW on Intel and remove non-standard wormhole, and it will be supported with non-Intel platforms. See bug 1762409

Comment 1

[Yury Delendik (:yury)]

Attachment: Bug 1762413 - Optimize emscripten's sequence for _mm_maddubs_epi16. r?lth

Comment 2

[Yury Delendik (:yury)]

Try push https://treeherder.mozilla.org/jobs?repo=try&revision=36aabb0d60b6cb27021028431687ce596b1e82be

Just to notice:

• the sequence folding will be available for regular content (not only extensions)
• is not available on ARM64, but functionality will not be lost there, it will be somewhat 9x slower
• the bergamot team is advised to add testing so it verifies the exact PMADDUBSW-sequence is generated when wasm is built.

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/8aa921dbad44
Optimize emscripten's sequence for _mm_maddubs_epi16. r=lth

Comment 4

[Cristian Tuns]

https://hg.mozilla.org/mozilla-central/rev/8aa921dbad44