Can bypass validation of `asm!` in a function that is never used via `inline(always)`

[japaric]

STR

NOTE target = x86_64-unknown-linux-gnu in all cases

This is rejected:

#![feature(asm)]

unsafe fn bkpt() {
    // ARM specific instruction
    asm!("bkpt");
}

$ rustc asm.rs
error: <inline asm>:1:2: error: invalid instruction mnemonic 'bkpt'
        bkpt
        ^

But this is accepted:

#![feature(asm)]

#[inline(always)]
unsafe fn bkpt() {
    asm!("bkpt");
}

$ rustc asm.rs && echo OK
OK

Don't worry though, because you can't actually use the ARCH specific instruction:

#![feature(asm)]

fn main() {
    unsafe {
        bkpt();
    }
}

#[inline(always)]
unsafe fn bkpt() {
    asm!("bkpt");
}

$ rustc asm.rs
error: <inline asm>:1:2: error: invalid instruction mnemonic 'bkpt'
        bkpt
        ^

Additional comments

This is a pretty useful bug/feature! You can use it cargo test the arch-agnostic parts of a crate
designed for ARM on x86 without adding a bunch of cfg attributes:

#![feature(asm)]

// ARM only
#[inline(always)]
pub unsafe fn control() -> u32 {
    let r: u32;
    asm!("msr CONTROL, $0" : "=r"(r));
    r
}

// arch-agnostic stuff
pub fn foo() -> bool {
    true
}

#[test]
fn bar() {
    assert!(foo());
}

$ cargo test
running 1 test
test bar ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured

Without the inline(always) trick, you'll have to add a cfg(target_arch = "arm") to the bkpt
function to get this to compile.

I don't think we want people to rely on this though. Specially, if the inline(always) trick stops
working in the future.

Meta

$ rustc -V
rustc 1.13.0-nightly (4f9812a59 2016-09-21)

cc @brson I'll let you decide how bad this is.

[hanna-kruppe]

This is probably a side effect #36524 (which avoids generating code for #[inline] functions in crates/CGUs that don't call them) combined with this asm validation being performed by LLVM (i.e., after codegen).

[japaric]

Can confirm that the inline(always) variant is rejected by nightly-2016-09-20 which doesn't include #36524. cc @michaelwoerister

[alexcrichton]

cc #29722, the tracking issue for asm!

[michaelwoerister]

Good catch, @japaric!
Yes, if asm! statements are only verified in trans, than my recent PR will prevent the compiler from doing so in many cases. Two questions:

• Will asm be verified later on, once the inline function gets instantiated somewhere?
• How much is asm verification bound to trans? Could it also be done when just the MIR is available?

[michaelwoerister]

OK, it seems that

• @japaric already answered the first question in the initial description. Verification happens once the function is used somewhere, and
• this verification is done by LLVM, so there is no simple way to do it without generating LLVM IR for the given function.

The question is, how do we want to handle this. I'd very much prefer if translation would always succeed once things pass analysis. On the other hand, before the PR, the compiler also failed as late as translation, so the only difference is that now things can fail in a downstream crate. I bet though that even before the inlining PR, we had exactly the same problem with generic functions, that is, an asm! block would not get verified either until the generic function was instantiated somewhere.

[michaelwoerister]

Nominating because I want to discuss if there is anything we can do or if we even want to do something here. The specific question is:

• Inline assembly only gets fully verified once it ends up in LLVM, which is not the case for unused generic functions and now also inline functions. Are we OK with seemingly successfully compiling code into an rlib that will fail in a downstream crate because it contains invalid inline asssembly?

[pnkfelix]

marking P-low

as noted above, this is another special case instance of code that can end up in an rlib and then cause a compilation failure further downstream. (We cannot really do anything about it without making our own verification of asm, rather than relying on LLVM's)