When an integer is converted into floating point in subword vector extract,
it can be done in 2 instructions instead of the 3+ instructions it generates
right now. This patch removes the uncessary generation.
Differential: https://reviews.llvm.org/D100604