webgl: populate uniformLocsById on first glUniform

[vittorioromeo]

PR created with AI assistance.

$webglGetUniformLocation reads program.uniformLocsById to resolve a uniform location integer to a WebGLUniformLocation. uniformLocsById is reset to 0 by glLinkProgram and lazily allocated by
$webglPrepareUniformLocationsBeforeFirstUse, but that prepare call only ran on the glGetUniformLocation() path. As a result, calling glUniform*() on a freshly linked program before any glGetUniformLocation() — which is the intended usage with -sGL_EXPLICIT_UNIFORM_LOCATION and explicit layout(location=N) — silently no-opped: $webglGetUniformLocation read uniformLocsById[location] off the integer 0, returned undefined, and GLctx.uniform*() dropped the call with no error.

Fix: call $webglPrepareUniformLocationsBeforeFirstUse(p) at the top of $webglGetUniformLocation, inside the GL_TRACK_ERRORS guard. The prepare call is idempotent (guards on !uniformLocsById), so the steady-state overhead is a single property read.

Adds browser.test_webgl_uniform_before_get_location with two parameterizations (default and -sGL_ASSERTIONS).

Fixes: #26672

[juj]

Can you gate this block with

#if GL_EXPLICIT_UNIFORM_LOCATION
   // ...
   webglPrepareUniformLocationsBeforeFirstUse(p);
#endif

so that WebGL programs that do not use the explicit uniform location extension, won't need to take this micro-perf-hit.

[vittorioromeo]

Makes sense, will do.

[juj]

Maybe here we might want to test layout(location = 6) or something else arbitrary, since default 0 could theoretically pass by the virtue of "the only uniform gets assigned an ID 0" kind of a behavior?

[vittorioromeo]

Makes sense, will do.

[juj]

Thanks. Looks solid, one comment on the perf/code size aspect of the feature.

[vittorioromeo]

@juj: addressed both comments, thanks for the review.

[juj]

Nice, thanks!

[vittorioromeo]

@juj: any idea why the CI is failing? The tests pass when I try them locally via runner.py

[juj]

@juj: any idea why the CI is failing? The tests pass when I try them locally via runner.py

It is the _assertions variant of the test that is failing. If you locally run

test/runner browser.test_webgl_uniform_before_get_location_assertions

you should see the failure.

This error is a real issue with the ASSERTIONS logic. The assertion code is too early with respect to the lazily initialized uniforms code, when GL_EXPLICIT_UNIFORM_LOCATION is being used.

To fix, it's probably simplest to add the following diff:

diff --git a/src/lib/libwebgl.js b/src/lib/libwebgl.js
index 549450fc92..a127f37d27 100644
--- a/src/lib/libwebgl.js
+++ b/src/lib/libwebgl.js
@@ -603,7 +603,9 @@ for (/**@suppress{duplicate}*/var i = 0; i <= {{{ GL_POOL_TEMP_BUFFERS_SIZE }}};

 #if GL_ASSERTIONS
     validateGLObjectID: (objectHandleArray, objectID, callerFunctionName, objectReadableType) => {
-      if (objectID != 0) {
+      // objectHandleArray may be uninitialized when GL uniforms are lazily initialized, and glUniform* is called
+      // for the first time before uniforms have been populated. So ignore this validation if the handle array is not present.
+      if (objectID != 0 && objectHandleArray) {
         if (objectHandleArray[objectID] === null) {
           err(`${callerFunctionName} called with an already deleted ${objectReadableType} ID ${objectID}!`);
         } else if (!(objectID in objectHandleArray)) {

because it looks like refactoring the assertions here so that all the GL.validateGLObjectID(GLctx.currentProgram.uniformLocsById, location, ...); checks occur after the lazy init has been done, can get quite hairy, since there are multiple conditional paths depending on the WebGL entry point type.

[vittorioromeo]

@juj: I see, my apologies. I assumed that./test/runner.py browser.test_webgl_uniform_before_get_location would run both variants of the parametrized test automatically.

I applied your suggested diff to the PR.

[juj]

Great, thanks!