The big difference between the even and odd numbered shaders/glsl-array-bounds-* seems to be how the invalid index comes about.

In the odd-numbered ones, it comes directly from shader code, e.g.:

void main() {
  int idx = 20;
  array[idx]
}

In the even-numbered ones, it comes from a uniform that I guess is passed in:

uniform int idx;

void main() {
  array[idx]
}

(And the various tests check fragment/vertex shaders, positive/negative offsets, and the place where the idx comes from, so 8 tests total.)

Although I'm not sure there's actually a problem here... the program is doing something illegal. You get segfault notifications in dmesg too...

Is this still a bug with recent kernels and other software? Please retest with 3.19-rc4 as there were some fixes for these chips in particular.

Noted on IRC by pmoureau:

Tests {2-8} are working fine while {10, 12} are still failing with a similar error on MCP79/7A (the closest we got)

The piglit test got changed in the meanwhile. It's an out-of-bounds access, I think the error in dmesg is legit...ish. Not sure. We could turn off error reporting, which afaik is what the blob does.

We don't disable error reporting, so you'll see stuff like that in dmesg. Expected.