asm! can't let you use lateout("rbx") (LLVM needs it)

[noncombatant]

This is about an unstable feature, so maybe you don't consider that a regression, exactly. This seemed like the best bug template to me; I hope it's OK.

I'm reading the asm! document at https://doc.rust-lang.org/beta/unstable-book/library-features/asm.html. The cpuid example has this code:

#![allow(unused)]
#![feature(asm)]
fn main() {
let ebx: u32;
let ecx: u32;

unsafe {
    asm!(
        "cpuid",
        // EAX 4 selects the "Deterministic Cache Parameters" CPUID leaf
        inout("eax") 4 => _,
        // ECX 0 selects the L0 cache information.
        inout("ecx") 0 => ecx,
        lateout("ebx") ebx,
        lateout("edx") _,
    );
}

println!(
    "L1 Cache: {}",
    ((ebx >> 22) + 1) * (((ebx >> 12) & 0x3ff) + 1) * ((ebx & 0xfff) + 1) * (ecx + 1)
);
}

But when you press the Play button to try it out, you get:

   Compiling playground v0.0.1 (/playground)
error: invalid register `ebx`: rbx is used internally by LLVM and cannot be used as an operand for inline asm
  --> src/main.rs:15:9
   |
15 |         lateout("ebx") ebx,
   |         ^^^^^^^^^^^^^^^^^^

error: could not compile `playground` due to previous error

I hacked it up locally (building with Rust Nightly), and this seems to work:

    unsafe {
        asm!(
            "cpuid",
            // EAX 4 selects the "Deterministic Cache Parameters" CPUID leaf
            inout("eax") 4 => _,
            // ECX 0 selects the L0 cache information.
            inout("ecx") 0 => ecx,
            //lateout("ebx") ebx,
            lateout("edx") _,
        );
        asm!(
                "mov {:e}, ebx", out(reg) ebx
        );
    }

I.e. I comment out the lateout("ebx") line, and add a second macro invocation to get the value of ebx.

But I assume there is a better way. :) The docs should be updated to show the better way. Thanks!

[JohnTitor]

It's allow_fail at first (3233565) so not a regression, we should just update the doc.

[asquared31415]

Isn't it still incorrect to call cpuid then? If it does clobber ebx then we're also just not telling llvm that we're breaking it. I think the correct code saves and restores ebx around the cpuid code, and doesn't use explicit register operands, instead getting generic out(reg) and moving to those registers from the cpuid output registers.

[asquared31415]

Here is a working example of code that does not clobber rbx by saving it.

Additionally, I opted to use CPUID 0 because 4 is not supported on AMD processors whereas 0 should be universal to all x86 processors that implement CPUID. It's reasonable that users could copy this code, and it should report a more reasonable value for more users.

[asquared31415]

I could make the pr to correct this if nobody has any issues with that change

[JohnTitor]

@asquared31415 Feel free to go ahead :)

[hkBst]

The unstable book does not seem to contain this anymore, so perhaps this can be closed?

[Amanieu]

Yes this has been solved in the docs.