1132920 - SIMD x86: use SSE3 movddup / AVX vmovddup for shuffle pattern (0, 1, 0, 1)

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Benjamin Bouvier [:bbouvier] (inactive)]

movddup, aka Move One Double-FP and Duplicate, corresponds to the shuffle pattern (1, 2, 1, 2), so we could probably optimize it in the JITs.

Comment 1

[Benjamin Bouvier [:bbouvier] (inactive)]

Oops, the shuffle pattern is actually (0, 1, 0, 1), and it's already optimized with vmovlhps, which is SSE2, hence more spread.  My bad.  (Perhaps it could be useful for the float64x2 shuffle pattern (0, 0), though)

Comment 2

[Dan Gohman [:sunfish]]

movddup wouldn't require a the inputCopy, so it would actually be slightly nicer, at least on pre-AVX machines (or AVX machines too, until we re-enable AVX (bug 1124751)).

Comment 3

[Benjamin Bouvier [:bbouvier] (inactive)]

Attachment: movddup.patch

The use of twoByteOpSimdFlags is unfortunate, because it doesn't seem to relate to flags in any ways, but that seems to be the one variant that fits. Should I rename it, or add a comment that contrarily to the name, it doesn't affect flags?

Comment 4

[Dan Gohman [:sunfish]]

Comment on attachment 8571973 [details] [diff] [review]
movddup.patch

Review of attachment 8571973 [details] [diff] [review]:
-----------------------------------------------------------------

The way this kind of encoding is handled elsewhere is to use the regular twoByteOpSimd and pass invalid_xmm as the src0. However, if you want to rename twoByteOpSimdFlags and use that instead, I'm cool with that too. But do rename it if you do :-).

Comment 5

[Benjamin Bouvier [:bbouvier] (inactive)]

I've slightly changed the condition, so that we emit movddup if we have SSE3 but not AVX, to follow the remark in comment 2.

https://hg.mozilla.org/integration/mozilla-inbound/rev/73bddae4f266

Comment 6

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/73bddae4f266