Improve the code generated for construction of a 64-bit floating point
number from two 32-bit integers.

Previously, this moved FP->core, inserted, then moved core->FP for each
half. Now, we construct the double in an X register and move core->FP.
Typically, the temporary register aliases the input register, so the
sequence improves from six to two instructions.

Patch from Martyn Capewell <m.m.capewell@googlemail.com>.

BUG=

Review URL: https://codereview.chromium.org/1008003004

Cr-Commit-Position: refs/heads/master@{#27227}