Replacing e with (<must be false> ? ... : e) changes image dramatically

[afd]

[Shaders, screenshots, error log, reproduction steps, etc.] [ZIP]

This fragment shader:

https://github.com/mc-imperial/shader-compiler-bugs/blob/master/ARM-Mali-T628-Chromebook-WebGL/wrong_images/small-v100-webgl-804943b2b630b741_inv_variant_4/original.frag

leads to this image being rendered on one of our Intel GPUs, and a visually identical image being rendered on other GPUs we have tested:

On the Chromebook, using WebGL, the following image is rendered instead:

This fragment shader:

https://github.com/mc-imperial/shader-compiler-bugs/blob/master/ARM-Mali-T628-Chromebook-WebGL/wrong_images/small-v100-webgl-804943b2b630b741_inv_variant_4/variant.frag

is identical to the original fragment shader, except that a reference to variable k has been changed to (injectionSwitch.x > injectionSwitch.y ? 1.0 : k). This should have no effect on rendering if we set injectionSwitch to the vector (0.0, 1.0), because the condition in the ternary will evaluate to false.

When we set injectionSwitch to this value and render on Mali using the modified shader, we get this image:

which is visually identical to the image that we see rendered on other GPUs we tried.

We speculate that a compiler bug is being triggered when the original shader is used for rendering, and that our modification somehow masks the compiler bug.

[afd]

Reported to ARM here:

https://community.arm.com/message/39845

[afd]

Closing this issue as, like #26, ARM reported it to be independently fixed.