swift234

ArrayOpts: handle local variable arrays, not just inout args.

This improves RC4 at -Ofast by ~2.7x.

It also makes microbenchmarks more robust. Without this, we were
seeing dramatic performance artifacts due to inlining differences.

For example, if you remove the @inline attribute from Quicksort, this
will give us a 10x speedup.

Here are the >5% differences:

| benchmark     | base -O3 |  opt -O3 |  delta | speedup |
| RC4           |   746.00 |   686.00 |  60.00 |    8.7% |
| Histogram     |  1147.00 |  1086.00 |  61.00 |    5.6% |
| StringWalk    |  1334.00 |  1265.00 |  69.00 |    5.5% |
| SmallPT       |  1124.00 |  1069.00 |  55.00 |    5.1% |

(I think the non-RC4 results might be bogus)

| benchmark     |base-Ofast|opt -Ofast|   delta | speedup |
| RC4           |  1094.00 |   405.00 |  689.00 |  170.1% |
| EditDistance  |  1727.00 |  1396.00 |  331.00 |   23.7% |
| StringWalk    |  1043.00 |   937.00 |  106.00 |   11.3% |
| Dictionary    |   930.00 |  1009.00 |   79.00 |   -7.8% |
| Phonebook     |  2270.00 |  2131.00 |  139.00 |    6.5% |
| StringBuilder |  1842.00 |  1723.00 |  119.00 |    6.9% |

RC4 is definitely real. It's always between 2.5-3x faster.

The others are very flaky.

Swift SVN r20445