cmov: use ASM `XOR` for equality comparisons

[brxken128]

This is how we originally did equality comparisons for x86, but for some reason #[inline(always)] caused it to break entirely.

I'd prefer to do the XOR within the ASM block as it's just more consistent, and I'll work on testing this out later today.

[tarcieri]

We should probably stop using #[inline(always)]: rust-lang/stdarch#306

[brxken128]

We should probably stop using #[inline(always)]: rust-lang/stdarch#306

And change to just #[inline]? I thought this was best practice anyway, as LLVM/whatever will decide the rest.

[tarcieri]

Our main usage of #[inline(always)] before was as an alternative to #[target_feature] (which mandates the annotated function must be unsafe), but this lead to a lot of hard-to-debug codegen-related issues.