Bug 214393 – [WebGL2] vector-dynamic-indexing is failing a subtest on iOS

• Status: RESOLVED
• Resolution: MOVED
• Component: WebGL

[Dean Jackson]

The test webgl/2.0.0/conformance2/glsl3/vector-dynamic-indexing.html is failing on iOS.

In particular, the part of the test that is:

test: Index an inout parameter passed to an user-defined function with an index with side effects

#version 300 es
precision mediump float;
out vec4 my_FragColor;
uniform int u_zero;
int sideEffectCounter = 0;
void foo(inout float f) {
    float g = f + 2.5;
    modf(g, f);
}
int funcWithSideEffects() {
    sideEffectCounter++;
    return 2;
}
void main() {
    vec4 v = vec4(1.0, 2.0, 3.0, 4.0);
    foo(v[funcWithSideEffects()]);
    vec4 expected = vec4(1.0, 2.0, 5.0, 4.0);
    float f = 1.0 - distance(v, expected);
    if (sideEffectCounter != 1) {
        f = 0.0;
    }
    my_FragColor = vec4(1.0 - f, f, 0.0, 1.0);
}

It produces a red output, which indicates that the f value is always 0. It should be 1.

[Radar WebKit Bug Importer]

<rdar://problem/65641563>

[Dean Jackson]

This also fails on macOS 11 (Apple Silicon)

[Dean Jackson]

I believe the sideEffectCounter variable is ending up with a value of 2 somehow.

[Dean Jackson]

Yes. It gets set to 2. Weird.

The translated shader source (on macOS apple silicon) is:

#version 410

out vec4 _umy_FragColor;
uniform int _uu_zero;
int _usideEffectCounter = 0;

void _ufoo(inout float _uf){
  float _ug = (_uf + 2.5);
  modf(_ug, _uf);
}

int _ufuncWithSideEffects(){
  (_usideEffectCounter++);
  return 2;
}

void main(){
  (_umy_FragColor = vec4(0.0, 0.0, 0.0, 0.0));
  vec4 _uv = vec4(1.0, 2.0, 3.0, 4.0);
  _ufoo(_uv[int(clamp(float(_ufuncWithSideEffects()), 0.0, float(3)))]);
  vec4 _uexpected = vec4(1.0, 2.0, 5.0, 4.0);
  float _uf = (1.0 - distance(_uv, _uexpected));
  if ((_usideEffectCounter != 1))
  {
    (_uf = 0.0);
  }
  (_umy_FragColor = vec4((1.0 - _uf), _uf, 0.0, 1.0));
}

But even in this form the funcWithSideEffects() is only called once. For some reason it is being called twice in the iOS GL implementation.

[Kenneth Russell]

Sounds like a bug in the GLSL compiler, perhaps in the implementation of the clamp operator when it's expanding out the AST. It sounds like it's duplicating the AST node which makes that call, rather than assigning its value to a temporary and referencing that temporary multiple times.

Could you try modifying your ANGLE build to pass SH_CLAMP_WITH_USER_DEFINED_INT_CLAMP_FUNCTION for:
  ShArrayIndexClampingStrategy ArrayIndexClampingStrategy
when passing the ShBuiltInResources to ANGLE's internal GLSL compiler instance? Does that work around the bug?

[Kimmo Kinnunen]

Currently fixed in specific OpenGL driver versions.
Currently iOS will use Metal, which does not have this bug.
Will update the expectation in bug 223934

[Kimmo Kinnunen]

Created attachment 437406 [details]
Patch

[Brent Fulgham]

The fix for this issue was needed outside the WebKit project, therefore this is being resolved as 'Moved'.

This should now be fixed in shipping software.