glibc ld.so has an optimization when resolving a symbol that checks
whether or not the upper 128 bits of the ymm registers are zero. If
so it uses "cheaper" instructions to save/restore them using the xmm
registers. If those upper 128 bits contain undefined values memcheck
will issue an Conditional jump or move depends on uninitialised value(s)
warning whenever trying to resolve a symbol.

https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/x86_64/dl-trampoline.h#l53

Apparently it is cheaper to check the upper bits are set than to invoke AVX unnecessarily (it is more energy efficient if you only use sse instructions on some CPUs so the cores don't have to "power up" the full AVX engine).

This triggers in our sh-mem-vecxxx test cases because we there explicitly make sure those bits are (partially) undefined.

We can workaround it my just resolving all symbols early. Which is what my first patch does.

But Tom Hughes said it might be better to make this a generic default suppression. I haven't seen this trigger outside the testsuite. But I guess that does make sense. That is patch number 2.