1096684 - Optimize SIMD shuffle and swizzle sequences with SSE3 and SSE4.1

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement)

Description

[Dan Gohman [:sunfish]]

SSE3 and SSE4.1 provide movsldup, movshdup, insertps, blendps, and pextrd, which are advantageous in SIMD shuffle, insert, and extract sequences.

Comment 1

[Dan Gohman [:sunfish]]

Attachment: extraname.patch

This patch implements extraName() for the LIR extract and insert element classes, for more readable debug output.

Comment 2

[Dan Gohman [:sunfish]]

Attachment: insertps-blendps.patch

This patch optimizes shuffles with insertps and blendps.

Comment 3

[Dan Gohman [:sunfish]]

Attachment: pextrd.patch

This patch adds the pextrd instruction and optimizes extract-element and boxDouble on x86 (where unboxDouble already uses pinsrd).

Comment 4

[Dan Gohman [:sunfish]]

Attachment: movsldup-movshdup.patch

This patch adds movsldup and movshdup and uses them for swizzles.

Comment 5

[Benjamin Bouvier [:bbouvier]]

Comment on attachment 8520298 [details] [diff] [review]
extraname.patch

Review of attachment 8520298 [details] [diff] [review]:
-----------------------------------------------------------------

Cool.

::: js/src/jit/LIR-Common.h
@@ +162,5 @@
> +          case LaneX: return "lane x";
> +          case LaneY: return "lane y";
> +          case LaneZ: return "lane z";
> +          case LaneW: return "lane w";
> +          default: return "unknown lane";

can you please remove the default case, such that the compiler complains if one adds a member to SimdLane and forgets to add the lane name here?

@@ +211,5 @@
> +          case LaneX: return "lane x";
> +          case LaneY: return "lane y";
> +          case LaneZ: return "lane z";
> +          case LaneW: return "lane w";
> +          default: return "unknown lane";

ditto

Comment 6

[Benjamin Bouvier [:bbouvier]]

Comment on attachment 8520301 [details] [diff] [review]
insertps-blendps.patch

Review of attachment 8520301 [details] [diff] [review]:
-----------------------------------------------------------------

Looks good, r+ with tests.

::: js/src/jit/shared/BaseAssembler-x86-shared.h
@@ +379,5 @@
>  
>      typedef enum {
>          OP3_ROUNDSS_VsdWsd  = 0x0A,
>          OP3_ROUNDSD_VsdWsd  = 0x0B,
> +        OP3_BLENDPS_VpdWpdIb = 0x0C,

nit: should be OP3_BLENDPS_VpsWpsIb as (legit) operands are packed singles

::: js/src/jit/shared/CodeGenerator-x86-shared.cpp
@@ +2476,5 @@
>      MOZ_ASSERT(numLanesFromLHS < 4);
>  
> +    // If all values stay in their lane, this is a blend.
> +    if (AssemblerX86Shared::HasSSE41()) {
> +        if (x % 4 == 0 && y % 4 == 1 && z % 4 == 2 && w % 4 == 3) {

Could you add tests in jit-tests/tests/asm.js/testSIMD.js for this case and the ones below, below the line commented as:

// Impl-specific special cases for shuffle (case and swapped case)

Comment 7

[Benjamin Bouvier [:bbouvier]]

Comment on attachment 8520303 [details] [diff] [review]
pextrd.patch

Review of attachment 8520303 [details] [diff] [review]:
-----------------------------------------------------------------

Nice!

::: js/src/jit/shared/BaseAssembler-x86-shared.h
@@ +3728,5 @@
>  
> +    void pextrd_irr(unsigned lane, XMMRegisterID src, RegisterID dst)
> +    {
> +        MOZ_ASSERT(lane < 4);
> +        spew("pextrd     $%u, %s, %s", lane, nameFPReg(src), nameIReg(4, dst));

could be $%x for symmetry with pinsrd_irr

@@ +3737,5 @@
> +
> +    void pextrd_imr(unsigned lane, XMMRegisterID src, int offset, RegisterID base)
> +    {
> +        MOZ_ASSERT(lane < 4);
> +        spew("pextrd     $%u, %s, %s0x%x(%s)", lane, nameFPReg(src),

ditto

Comment 8

[Benjamin Bouvier [:bbouvier]]

Comment on attachment 8520305 [details] [diff] [review]
movsldup-movshdup.patch

Review of attachment 8520305 [details] [diff] [review]:
-----------------------------------------------------------------

Cool!

::: js/src/jit/shared/CodeGenerator-x86-shared.cpp
@@ +2416,5 @@
>      uint32_t z = ins->laneZ();
>      uint32_t w = ins->laneW();
>  
> +    if (AssemblerX86Shared::HasSSE3()) {
> +        if (ins->lanesMatch(0, 0, 2, 2)) {

Can you add tests for these two specific patterns in testSIMD.js as well please?

Comment 9

[Dan Gohman [:sunfish]]

Nits addressed, and as noted on irc, the tests are in.

https://hg.mozilla.org/integration/mozilla-inbound/rev/9bd6e299c84a
https://hg.mozilla.org/integration/mozilla-inbound/rev/191a52db5011
https://hg.mozilla.org/integration/mozilla-inbound/rev/cf5beb0f5d25
https://hg.mozilla.org/integration/mozilla-inbound/rev/4c04203003c2

Comment 10

[Wes Kocher (:KWierso)]

https://hg.mozilla.org/mozilla-central/rev/9bd6e299c84a
https://hg.mozilla.org/mozilla-central/rev/191a52db5011
https://hg.mozilla.org/mozilla-central/rev/cf5beb0f5d25
https://hg.mozilla.org/mozilla-central/rev/4c04203003c2