Switch from llvm_asm!() to new asm!() macro

[coolreader18]

To be honest I'm not the most familiar with inline assembly, but afaict the new version is correct. The assembly output is identical, at least, checking RawRwLock::bump_shared_slow (which calls both elision impls) using cargo asm.

This should be ready for use when asm!() is stabilized, cf (well, your own stabilization report I guess) rust-lang/rust#72016 (comment)

[Amanieu]

This code is actually incorrect on the x32 ABI: you need to force the use of the :e modifier for both registers so that a 32-bit cmpxchg is used instead of a 64-bit one.

[coolreader18]

Not that I'd know more about this than you, but are you sure? The unstable book looks like it implies that eax will be chosen on x86 32-bit (https://doc.rust-lang.org/unstable-book/library-features/asm.html)

Architecture	Register class	Modifier	Example output	LLVM modifier
x86-32	reg	None	eax	k
x86-64	reg	None	rax	q

Is it just by chance that it's picking the correct e registers?

[Amanieu]

x32 is not the same as x86. It's the 32-bit ABI for x86_64 where it uses the x86_64 ISA but with 32-bit pointers. You can try it yourself with the x86_64-unknown-linux-gnux32 target.

[Amanieu]

You also need :e here for 32-bit pointer size.