miri ICE promote const: impossible case reached

[gnzlbg]

Playground:

#![feature(const_fn, rustc_attrs, attr_literals, untagged_unions)] 

/// A single unconstrained byte (0-255).
#[derive(Copy, Clone, Debug)]
pub struct ImmByte(u8);
impl ImmByte {
    /// Constructor
    #[inline]
    #[rustc_args_required_const(0)]
    pub const fn new(value: u8) -> Self {
        ImmByte(value)
    }
}

pub struct v128([u8; 16]);

impl v128 {
    #[inline]
    #[rustc_args_required_const(0)]
    pub const unsafe fn const_(imm: [ImmByte; 16]) -> v128 {
        union U {
            imm: [ImmByte; 16],
            vec: v128,
        }
        U { imm }.vec
    }
}

fn main() {
    let _a = v128::const_([
        ImmByte::new(0),
        ImmByte::new(1),
        ImmByte::new(2),
        ImmByte::new(3),
        ImmByte::new(4),
        ImmByte::new(5),
        ImmByte::new(6),
        ImmByte::new(7),
        ImmByte::new(8),
        ImmByte::new(9),
        ImmByte::new(10),
        ImmByte::new(11),
        ImmByte::new(12),
        ImmByte::new(13),
        ImmByte::new(14),
        ImmByte::new(15),
    ]);
}

produces

error: internal compiler error: librustc_mir/transform/promote_consts.rs:334: impossible case reached

thread 'main' panicked at 'Box<Any>', librustc_errors/lib.rs:578:9
note: Run with `RUST_BACKTRACE=1` for a backtrace.
error: aborting due to previous error

cc @oli-obk

[cramertj]

Not that this fixes the issue, but [ImmByte; 16] and v128([u8; 16]) aren't necessarily guaranteed to have the same representation without #[repr(C)] or #[repr(transparent)] on both v128 and ImmByte.

[oli-obk]

minified:

#![feature(const_fn, rustc_attrs, attr_literals)] 

#[rustc_args_required_const(0)]
pub const fn new(value: u8) -> u8 {
    value
}

#[rustc_args_required_const(0)]
pub unsafe fn const_(_: u8) {
    unimplemented!()
}

fn main() {
    let _a = const_(new(0));
}

https://play.rust-lang.org/?gist=2032873a12df78fc3e4537d26735a050&version=nightly&mode=debug&edition=2015

@eddyb is this a victim of the backwards algorithm for promotion? Just a hunch, but it might have promoted away the call to new and then end up trying to promote the 0, but it's not in the place where it was expected, because it now lives inside the promoted. So it ends up trying to promote an argument in a basic block that was already promoted away

[gnzlbg]

thanks @cramertj , I'll update the PR that triggered this - for context rust-lang/stdarch#549

[oli-obk]

As a workaround until we fix this ICE, you can introduce an intermediate explicit constant for the argument to v128::const_. Or is this part of a user-exposed API?

[gnzlbg]

This is part of a user-exposed API, but that workaround should allow me to at least get the const tests running.

[eddyb]

Just a hunch, but it might have promoted away the call to new and then end up trying to promote the 0, but it's not in the place where it was expected, because it now lives inside the promoted.

It won't remove the definition if there's more than one use (unless this is fallout from your changes?)

[oli-obk]

There's only one use. I don't see how my changes could have caused this, but it seems possible. I'll do a bisection

[eddyb]

Oh, I see, #[rustc_args_required_const] is Candidate::Argument and that's missing a check like:

rust/src/librustc_mir/transform/promote_consts.rs

Lines 376 to 379 in a3ad012

	if temps[local] == TempState::PromotedOut {
	// Already promoted.
	continue;
	}

[oli-obk]

Fix instructions:

• Add a IndexVec<BasicBlock, bool> which, like the temps variable, is used to mark blocks whose terminator has been promoted out.
• skip promoting all such terminators when encountering said block id in a Candidate::Argument (

  rust/src/librustc_mir/transform/promote_consts.rs

  Line 384 in a3ad012

  Candidate::Argument { .. } => {}

  )

[eddyb]

@oli-obk You can use a BitVector<BasicBool>, although I'm not even sure it's necessary.
A promoted call ends up as a Goto, so you could just skip it based on that?