On Intel, xchg stalls the pipeline. We use xchg to implement swap
moves. In a separate exploration, the presence of xchg in a very hot
loop, due to a change in register allocation, lead to over 20% regression.

Simply changing that instruction with push/mov/pop (almost) eliminated
the regression.

In light of that, I removed uses of xchg. This leads to more instructions,
though. That is particularly problematic for long cycles, which, today,
we translate to successions of swaps.

I plan to address this cycle issue in a separate change. For now, the
goal is to unblock the initial work that lead here.

Review URL: https://codereview.chromium.org/1580233003

Cr-Commit-Position: refs/heads/master@{#33282}