Functions that read from out parameters have undefined behavior

[magcius]

void foo(out float a) {
  a += 2.0f;
}

void main() {
  float x = 5.0f;
  foo(x);
  gl_FragColor = vec4(x);
}

This is not an error, and is undefined behavior, but it works fine on some GPUs because their drivers effectively treat out as inout. However, I've encountered issues with other GPU drivers that are more strict. While we probably can't make it an error due to backcompat, maybe a compiler warning should be emitted for this?

[lexaknyazev]

Given the WebGL robustness restrictions, this should be behave like the following everywhere

void foo(out float a) {
  float _a = 0.0;
  _a += 2.0f;
  a = _a;
}

void main() {
  float x = 5.0f;
  foo(x);
  // x == 2.0
  gl_FragColor = vec4(x);
}

Please specify what do you see on which platforms.

[magcius]

I saw the behavior on Firefox and Linux. mesa treated out variables as undefined contents, and I suspect that Firefox is passing shaders through to the system GL implementation as-is.

[kdashg]

You can use WEBGL_debug_shaders.getTranslatedShaderSource() to see what we (and/or other browsers) are passing to the driver.
We do not pass the driver shaders as-is, though it's possible we're mistakenly missing a safety transform that e.g. Chromium is not.

[magcius]

https://jsbin.com/fiqazoqula/edit?js,output

Tested with Firefox 123.0 on Windows 10, here's what I see as the output. Firefox is generating code that has undefined behavior.

void webgl_550c7d686f02ef30(out highp float webgl_cc4b575f1e53d03d){
  const highp float s15e9 = 0.2;
  (webgl_cc4b575f1e53d03d += s15e9);
}
void main(){
  (gl_FragColor = vec4(0.0, 0.0, 0.0, 0.0));
  highp float webgl_350702f5a0424893 = 0.5;
  webgl_550c7d686f02ef30(webgl_350702f5a0424893);
  (gl_FragColor = vec4(webgl_350702f5a0424893));
}

I don't have access to a Linux machine currently, so I can't test Chrome on OpenGL.

Chrome 124.0.6330.0 on ANGLE D3D11 changes the parameter to be an inout, which does not match your expected robustness transform:

#define GL_USES_FRAG_COLOR
void f_test(inout float _v)
{
(_v += 0.2);
}
@@ PIXEL OUTPUT @@

PS_OUTPUT main(@@ PIXEL MAIN PARAMETERS @@){
@@ MAIN PROLOGUE @@
float _x5628 = {0.5};
f_test(_x5628);
(gl_Color[0] = vec4_ctor(_x5628));
return generateOutput();
}

[magcius]

A friend ran the script on Chrome 120.0.6099.199 on Linux/OpenGL, which provided the following results:

#version 450
out vec4 webgl_FragColor;
void _utest(out float _uv){
  (_uv = 0.0);
  (_uv += 0.2);
}
void main(){
  (webgl_FragColor = vec4(0.0, 0.0, 0.0, 0.0));
  float _ux = 0.5;
  _utest(_ux);
  (webgl_FragColor = vec4(_ux));
}

So it appears ANGLE's OpenGL shader translator is applying the transform @lexaknyazev expected to see here.

[kenrussell]

@magcius what's the output of Firefox's about:support on the Windows 10 machine that generated [this output)(https://github.com//issues/3628#issuecomment-1972168018)? I thought Firefox was using ANGLE's D3D11 backend by default there, so it should have similar output to Chrome's on Windows.

[kdashg]

I will update the spec with the expected initialize out to 0 behavior.

Bugs should be filed against browsers that don't do this properly already.

[magcius]

I could upload the about:support log somewhere, but it's fairly large and I don't want to post it directly in this issue. My (perhaps out of date) understanding is that Firefox doesn't use ANGLE's D3D11 backend for WebGL 2, it uses system OpenGL. I believe it uses ANGLE for shader translation, though.

[kenrussell]

Could you perhaps upload it to pastebin and post the link here? Or if you'd rather not have the info linked here, feel free to email the link to me at kbr at chromium dot org.

[kdashg]

Firefox has always used ANGLE's ES3->D3D11 implementation as our "GLES" driver on Windows for webgl2, at least by default. We have prefs to use native GL, but we've never "shipped" that config. This is however different from Chrome, in that Chrome effectively has ANGLE map directly webgl2->d3d11 on Windows, whereas Firefox uses ANGLE as a webgl-agnostic GLES driver, and does the webgl->gl(es) implementation within Gecko.

[kdashg]

For example, this machine's about:support reads:

WebGL 2 Driver Renderer:    Google Inc. (Intel) -- ANGLE (Intel, Intel(R) Arc(TM) A770 Graphics Direct3D11 vs_5_0 ps_5_0, D3D11-31.0.101.5186)
WebGL 2 Driver Version:     OpenGL ES 3.0.0 (ANGLE 2.1.19736 git hash: ddaf44ac75d5)

[kenrussell]

Per WebGL concall of 2024-03-21: some targeted tests are needed in the WebGL CTS (conformance/glsl/misc/?) for these cases.