Zero out only the first pointer when canceling cleanups, and don't zero out PODs

[pcwalton]

There's generally no need to zero out PODs, and no need to zero out anything beyond the first pointer of a non-POD data structure containing unique values. (Drop glue can simply bail out at the first null pointer it finds.)

LLVM does not optimize out redundant zeroing out of structs, unless all functions that operate on the struct are inlined and SROA happens. Neither does GCC. I believe (but haven't thought too deeply about it) that this is probably an unsafe optimization with the semantics of C. So this could be a rather large win for code size.

[graydon]

was this fixed? (is it part of the thing triggering valgrind problems currently?)

[bblum]

This optimization would be subsumed by #5016, but would be significantly easier than that project, so might get done on a faster timescale.

[thestinger]

This is still an issue, we do many zeroes of the entire value.

[pnkfelix]

closing; this problem will go away with the dtor reworking

[nikomatsakis]

Subsumed by #5016