Avoid redundant instructions in x86 inline asm

[sunfishcode]

This is inspired by an idea from @m-ou-se in #project-inline-asm in the Rust zulip. We can't use esi as an explicit inline asm operand, but we can request a register and test whether we got esi using .ifnes. When we do get esi, avoid swapping it into place.

And, eliminate a redundant push/pop in indirect_syscall5.

While here, add some comments and tidy up the arch/inline/mod.rs file.

[Urgau]

There are some ongoing debate in T-lang about what subsets of assembly instructions are guaranteed to be available and some that they want to disallow (because a different codegen or implementation might not have it), this includes the .if/.ifnes/... operands.

[sunfishcode]

There are some ongoing debate in T-lang about what subsets of assembly instructions are guaranteed to be available and some that they want to disallow (because a different codegen or implementation might not have it), this includes the .if/.ifnes/... operands.

Interesting, thanks for mentioning that! Restricting the set of directives is a good idea, and I encourage T-lang to do so; there are a lot of directives and they can interact with compilers in undocumented and undesigned ways.

If .if/.endif/etc. are disallowed, we can simply remove those directives here once nightly starts enforcing the new rules. They're a very minor optimization in this context.

[sunfishcode]

Ah, looking at the list here, it appears the plan is not to enforce the restrictions, but to say that things not on the list have assembler-specific behavior. Accordingly, I've now removed the optimization that uses .ifnes from this PR, to avoid depending on what will be considered assembler-specific behavior.