Precision of constant variables is dropped during operation.

[themikelester]

We have a case where a constant variable in an operation is being replaced with it's literal representation, dropping the precision qualifier. This variable was the only highp value in the operation, which means that the whole operation is now performed at mediump.

Here's the case (note that gl_FragCoord is mediump):

highp float time = 1.0;
highp float dither = dot( vec2( 171.0, 231.0 ), gl_FragCoord.xy + vec2( time ) );

is optimized to:

highp float tmpvar_17;
tmpvar_17 = dot (vec2(171.0, 231.0), (gl_FragCoord.xy + vec2(1.0, 1.0)));

The tmpvar_17 calculation now contains only one variable with precision, mediump gl_FragCoord, so the calculation is performed at mediump. Replacing the 1.0 literals with a highp float initialized to 1.0 fixes the issue. The bug here seems to be that constants are being replaced by literals without considering precision.