This patch adds a pass which expands vector insert/extract element with
a non-constant index into memory operations.

SIMD.js doesn't yet have vector-vector shifts or insert/extract by
non-constant index, so score those as expensive.

LLVM's optimizer generates types like <3 x float> in some cases, so we
should support them. Also, this is a chance to make use of
loadXYZ/storeXYZ and friends.

…d language extension.

… emscripten #3088

SIMD values can't simply be coerced from "0", so build a string for an
undef of the appropriate type.

…f35dd5

Instcombine is currently very conservative about combining vector
shuffles because it is afraid of pessimizing code translated into
generic form from xmmintrin.h intrinsics. For the current SIMD.js API,
it is more desirable to push JS engine code generators to emit good code
for the shuffles that users are actually doing, and less desirable to
preserve broken-up shuffles in their original form.