setjmp and its variants save the non-volatile GPRs and FPRs, but not the
non-volatile vector registers v20-v31, so after a longjmp they hold
whatever the code that jumped left in them. Julia runs into this when it
switches tasks with sigsetjmp/siglongjmp: the compiler keeps vector values
in v20-v31 and they come back from another task.
jmp_buf has had room for them all along. Add variants that save them and
pick those with ifuncs when the CPU has Altivec, so the same libc still
works on CPUs without it, like the e5500.
rtld can't use ifuncs, so it takes the variant for the CPU it's built
for. It needs the registers saved too: when it restarts a symbol lookup
to upgrade its lock, it longjmps out of functions that use v31, and a
lazily bound call came back with v31 zeroed.
MFC after: 1 week