1023404 - [meta] SIMD backend: implement SIMD operations for x86/x64

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect, P5)

Description

[Benjamin Bouvier [:bbouvier] (inactive)]

This bug should serve as a meta-bug for referencing all patches implementing SIMD operations.

As discussed with involved people, instructions should be grouped whenever possible, when they have similar semantics:
- with{X,Y,Z,W}
- same thing for withFlag
- add / sub / mul / div can be grouped as they all are binary arithmetic instructions
- equal / greaterThan / lessThan / notEqual
- min / max
- shiftLeft, shiftRight, shiftRightLogical
- and, or, not, xor
- lane accessors and shuffle

We should be careful and write some assertions that we are in a supported platform (in this case, x86 and x64). I will add a few blocking bugs as well as referencing others.

Comment 1

[Benjamin Bouvier [:bbouvier] (inactive)]

Attachment: Helpers for SIMD types

Just splitted up the patch from bug 1021716, as it had nothing to do there.

Comment 2

[Benjamin Bouvier [:bbouvier] (inactive)]

Attachment: SimdDataSize

There are a few places in the code where we need to memcpy / memcmp SIMD values. Let's introduce a variable that holds the current size of SIMD datum to be used in these cases.

Also adds a few names for iongraph.

Comment 3

[Benjamin Bouvier [:bbouvier] (inactive)]

Attachment: Helpers for SIMD types

Carrying forward r+ from sunfish.

Diff:
- renames "arity" into "length"
- SIMD => Simd

Comment 4

[Dan Gohman [:sunfish]]

Comment on attachment 8463350 [details] [diff] [review]
SimdDataSize

Review of attachment 8463350 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jit/shared/Assembler-shared.h
@@ +28,5 @@
>  
> +static const uint32_t SimdDataSize = 4 * sizeof(int32_t);
> +static_assert(SimdDataSize >= 4 * sizeof(int32_t), "SIMD data should be able to contain int32x4");
> +static_assert(SimdDataSize >= 4 * sizeof(float), "SIMD data should be able to contain float32x4");
> +static_assert(SimdDataSize >= 2 * sizeof(double), "SIMD data should be able to contain float64x2");

Would it make sense to name this Simd128DataSize? Then these could be == rather than >= and we wouldn't have to worry about unused bytes for memcmp.

Comment 5

[Benjamin Bouvier [:bbouvier] (inactive)]

Comment on attachment 8463350 [details] [diff] [review]
SimdDataSize

Moved to bug 1019831 for consistency and dependency requirements.

Comment 6

[Benjamin Bouvier [:bbouvier] (inactive)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/3a58450f815a

Comment 7

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/3a58450f815a

Comment 8

[Benjamin Bouvier [:bbouvier] (inactive)]

This one bug was open to implement the codegen for all the SIMD operators proposed by SIMD.js on x86. We can consider it done, and this should be a sufficient starting point for the wasm SIMD implementation; codegen methods will remain in the x86-shared MacroAssembler files.