Performance regression when alignment was set to 1

[dylanmckay]

When #64 was fixed by changing all types to be byte aligned, a performance regression was introduced.

The upstream LLVM commit that caused this is r314179.

I discovered the regression when fixing the build in r314183.

If you look at r314183, you will see that we previously used a register when pushing 16-bit values onto the stack, whereas now we needlessly use an extra register.