1313512 - SIMD: Implement (un)signed conversions between integer x4 types

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect, P5)

Description

[Jukka Jylänki]

Attempting to import SIMD.Int32x4.fromUint32x4() function to asm.js, with

  var SIMD_Int32x4=global.SIMD.Int32x4;
  var SIMD_Int32x4_fromUint32x4=SIMD_Int32x4.fromUint32x4;

I'm getting an error

  TypeError: asm.js type error: 'fromUint32x4' is not an operation supported by the SIMD type

STR: http://clb.demon.fi/bugs/int32x4_from_uint32x4/test_simd_int32x4.html

Comment 1

[Jukka Jylänki]

Further down in the file, lines

  var SIMD_Uint32x4=global.SIMD.Uint32x4;
  var SIMD_Uint32x4_fromInt32x4=SIMD_Uint32x4.fromInt32x4;  

are giving a similar error

  TypeError: asm.js type error: 'fromInt32x4' is not an operation supported by the SIMD type

Comment 2

[Benjamin Bouvier [:bbouvier] (inactive)]

Spidermonkey doesn't implement SIMD.Int32x4.fromUint32x4 or SIMD.Uint32x4.fromInt32x4; in general, these can be supplanted by the bit cast variants, since a number can be seen as signed or unsigned. I just found out these functions belong in the SIMD.js spec though, with the imprecise conversion rule: if a lane input can't fit the destination type, an error is raised.

Can you use SIMD.Int32x4.fromUint32x4Bits and SIMD.Uint32x4.fromInt32x4Bits instead? (these are type coercion no-op (i.e. no generated code) in asm.js). Or do you need the imprecise conversion rule to apply here?

Comment 3

[Jukka Jylänki]

When compiling SSE code (the most important case), Emscripten is actually already using only the *Bits variants, but it was just importing all of the SIMD.js functions unconditionally, since we don't have an instruction level mechanism of tracking what needs to be imported in.

Removed these for now in Emscripten: https://github.com/kripken/emscripten/commit/4092210034fe79ecdff1b8f69744cf770bcecbc3 . The only scenario this would affect is if user codebases might be using the intrinsics explicitly via the emscripten/vector.h (https://github.com/kripken/emscripten/blob/master/system/include/emscripten/vector.h), but I don't think anyone currently is.

Comment 4

[Benjamin Bouvier [:bbouvier] (inactive)]

OK, thanks! Retargeting the bug summary then. I think that if the specification mandates it, we *could* implement it, but it's likely to be low priority unless it's used in real world contexts or added in wasm. CC'ing other SIMD.js people in case they know more.

Comment 5

[Jakob Stoklund Olesen [:jolesen]]

It hasn't been requested for WebAssembly to my knowledge. A trapping int <-> uint conversion is pretty easy to implement efficiently using the existing SIMD.js primitives, so I don't see a big need for providing it.

Comment 6

[Benjamin Bouvier [:bbouvier] (inactive)]

SIMD.js is being removed (bug 1416723). (This could be added to wasm, but that's not the right forum to discuss it)