Revert "[Impeller] fix line/polygon depth and GLES scissor state. (#56494)"

[chinmaygarde]

This reverts commit 950e240 in #56494.

This caused a regression in a customer application (@lyceel).

I am not sure I should revert this whole patch. The goldens don't change if I just revert the patches to render_pass_gles.cc. So while that part was untested in the original commit, the other patches should be good. The original commit message did mention two separate issues were being addressed.

I am working on reduced test case with just the scissor regression with the customer.

[chinmaygarde]

From the original commit message.

Fixed in the same PR because the golden actually demonstrates both problems with GLES, and luckily I found 2. when I noticed 1.

If I back out just the disabling of the scissor test, the goldens still look identical to Metal/VK. Am I missing something?

[jonahwilliams]

I don't see an obvious way for the stencil setting to leak out of the render pass. Even if we leave it set, then a subsequent render pass would immediately disable it via gl.Disable(GL_SCISSOR_TEST) (at least until a new stencil value was set).

Is this bug happening because the stencil state is leaking into the host context libimpeller is embedded into?

[jonahwilliams]

Would like to see a repro, otherwise we'd just need to reland this to unblock Android GLES

[chinmaygarde]

Yeah, working on a repro. Found this regression late in the day.

otherwise we'd just need to reland this to unblock Android GLES

Thats the thing. The goldens don't change even if I just revert just the removal of the scissor test toggle. So, I was wondering if you added it by mistake. Either way, its not tested.

But let's revisit this in the morning. The reduced test case ought to be simple enough. And I haven't reasoned about the embedder either. So it could be that.

[chinmaygarde]

Spoke about this in person and the regression may be because of #56597. Either way, closing this.