1690483 - SIMD optimization x64/x86: Better code for variable swizzle

Status: RESOLVED FIXED

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

Description

[Lars T Hansen [:lth]]

The variable swizzle on intel can use PSHUFB to shuffle the bytes but the mask vector must first be sanitized so that out-of-range lanes in the mask have the high bit set. Currently we use a compare-with-constant-and-POR to do this (and we don't even inline the constant load in the compare, sigh) but it's possible to do better by saturating-add'ing a constant into the mask: https://github.com/WebAssembly/simd/issues/68#issuecomment-470825324

For specific code generation, I'm not sure if it's better to (a) splat a byte value into scratch / load the constant into scratch, and add the mask to the scratch, or (b) to move the mask to scratch and add a constant from memory into scratch. Either way the mask register is not volatile.

Also see https://github.com/WebAssembly/simd/issues/93 for more discussion, probably worth reading although it ranges across a bunch of topics.

Comment 1

[Yury Delendik (:yury)]

Attachment: Bug 1690483 - Use saturating add for mask of SIMD swizzle. r?lth

Comment 2

[Yury Delendik (:yury)]

There is some 3-4% gain in local microbenchmark test.

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/ee2bc38e681e
Use saturating add for mask of SIMD swizzle. r=lth

Comment 4

[Noemi Erli[:noemi_erli]]

https://hg.mozilla.org/mozilla-central/rev/ee2bc38e681e