ARM: Avoid duplicate vmla when merging vmul and vadd

ARM: Avoid duplicate vmla when merging vmul and vadd

Avoid generating duplicate vmla instructions for vmul/vadd sequences where the
vmul has more than one use.

For example: function f(a, b, c) { return (a * b) + c + (a * b); }

Previously, this would produce a vmul for the subexpression (a * b), then vmla
for (a * b) + c, then vmla for (a * b) + [(a * b) + c].

Now it produces vmul, vadd, vadd, as expected.

BUG=

[ulan, 2014-06-16 11:14:47]

lgtm