naked functions

[folkertdev]

tracking issue: rust-lang/rust#90957
earlier version: #1153

Pending on rust-lang/rust#128004. Once merged, we'll move for stabliziation.

I'm working off of the original PR to the reference, written 2+ years ago when the reference was less strict. I suspect some refinement and editing will be needed here to satisfy the quality requirements of the spec, and update the text to accurately reflect the current state and the exact guarantees/requirements.

cc @bstrie @traviscross @Amanieu @bjorn3

[Amanieu]

This is not true, naked functions cannot be duplicated/merged, unlike asm!.

[folkertdev]

and reading this over, in this case it seems important that we do actually guarantee (and hence the compiler can assume) that the instructions in the naked_asm! invocation are exactly the ones that will be executed. or is that too strict?

[chorman0773]

I think that poses problems on SPIR-V or NVPTX targets. I seem to recall that the loader just inlines the functions at the Shader Assembly level.

[folkertdev]

@chorman0773 just to be sure, you're referring to Amanieu's comment right?

What we do in rustc (or well, we will once that PR makes it through the queue) is that a naked function

#[naked]
extern "C" fn foo() {}

is emitted as something similar to

core::arch::global_asm!( 
    "foo:",
    "ret"
);

extern "C" {
    fn foo();
}

and we're relying on the compiler (rustc, llvm) not duplicating that bit of global assembly. I suspect we already rely on that in general, because global assembly can define symbols today.

If you still think this is problematic, how could we verify the behavior of this new codegen strategy for the targets you list?

[Amanieu]

and reading this over, in this case it seems important that we do actually guarantee (and hence the compiler can assume) that the instructions in the naked_asm! invocation are exactly the ones that will be executed. or is that too strict?

No, they may not be the same instructions that are executed since you can still do runtime patching. We only guarantee that assembly code only appears once in the object file for the purposes of symbols.

[folkertdev]

so, something I just thought of: we do duplicate the assembly when monomorphizing generic naked functions. So this phrasing should be a bit more specific I think.