[tests] Regression test for EGL-glibc TLS conflict

[amtep]

bionic and glibc have different layouts for TLS space.
Since libEGL used a bionic slot directly (in inlined code),
libhybris's hooks didn't translate it properly and libEGL ended
up overwriting some unrelated thread-local values in glibc.

The problem only showed up when linking with libGLESv2 (which
pulls in libEGL), not when linking with libEGL directly. That's
why the test was added to test_glesv2.c

[thp]

Good, although this doesn't detect overwrites with zero, right? What about initializing the array with a hardcoded random set of values, or generate random values into it at runtime, and then copy the whole array to a local variable and then compare both arrays? This would detect overwrites with zero, but obviously will fail to detect overwrites with whatever random value was written to it (probability of 1 / 256).

Other than that, LGTM.

[amtep]

Yes, I hadn't thought of anything writing zeroes.
There are some technical issues with making it nonzero... the bug that inspired this test was something that happened during library init, so before main().
The slots could be initialized at declaration, though not with random values. That would put the array in .tdata section instead of .tbss but I think they are allocated to the TLS the same way.

[amtep]

I modified it so that the array is initialized with non-zero values.
It still detects the bug.

[thp]

LGTM.

[Thaodan]

Close?