glsl-1.20: Test various aspects of do-while loop scoping

This test also inadvertently tests the sequence operator. At the time this test was written, Mesa would not compile it. See also !569 (merged). glslangValidator also fails to compile it (see https://github.com/KhronosGroup/glslang/issues/2744), but NVIDIA's closed-source driver does compile it.

Once the compilation problem is resolved, I would like to wait to land this test until we are at least sure it won't lead to a GPU hang. The original version of this test would, but this version should be safe. Be making various tweaks to the test (adding { } around the body of the loop, s/int i/i/ inside the loop), I'm pretty confident that it tests everything the way I intend without causing an infinite loop.

added Tests label

requested review from @tarceri

mentioned in merge request !569 (merged)

Wouldn't we end up in the last else of the sequence operator was producing the wrong result? i.e isn't the last part of the comment in the wrong block?

Otherwise the test looks good. Thanks Ian.

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>

I tried several things to see various results.

I changed the loop to this, and I got green:

    do {
        int i = (j++, 0);
    } while (i <= 0 && j < 3);

I changed the loop to this, and I got blue:

    do
        i = j++;
    while (i <= 0 && j < 3);

This was the scenario I was trying to describe in that comment. The loop test uses the i from inside the loop, and that i is assigned j++ instead of 0.

I changed the loop to this, and I got red:

    do
        i = (j++, 0);
    while (i <= 0 && j < 3);

This is the case where the loop test uses the i from inside the loop, but everything else is working properly.

I just now changed the loop to this, and I got black:

    do
        i = 0;
    while (i <= 0 && j < 3);

Which indicates there's another bug. Setting INTEL_DEBUG=fs shows that the whole shader is optimized away. Or do we intentionally nuke shaders that we know have infinite loops?

Or do we intentionally nuke shaders that we know have infinite loops?

Yes I believe something was added for this a while ago.

Update: At least I thought there was, but I can't seem to find the commit.

Update2: Maybe this? mesa!2307 (merged)

Yeah, that looks like it would do it. Okay. This seems okay.

I'm going to add to the end of that comment, "The loop terminated when i was 1, and, due to post-increment, j is 2." How about that?

Right. Yeah that makes sense.

changed this line in version 2 of the diff

added 2 commits

• 99be1b06 - 1 commit from branch mesa:main
• 59e1c93d - glsl-1.20: Test various aspects of do-while loop scoping

Compare with previous version

assigned to @marge-bot

mentioned in commit idr/piglit@da5b4601

added 1 commit

• da5b4601 - glsl-1.20: Test various aspects of do-while loop scoping

Compare with previous version

merged