Support for `core::mem::forget_unsized`

[Jules-Bertholet]

#![feature(forget_unsized, unsized_fn_params)]

fn main() {
    let b: Box<[u32]> = Box::new([1, 2, 3]);
    std::mem::forget_unsized(*b);
}

Miri fails to run the above program with the following error:

error: unsupported operation: unsized locals are not supported
   --> /home/julesbertholet/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/mem/mod.rs:159:24
    |
159 |     intrinsics::forget(t)
    |                        ^ unsized locals are not supported
    |
    = help: this is likely not a bug in the program; it indicates that the program performed an operation that the interpreter does not support
    = note: BACKTRACE:
    = note: inside `std::mem::forget_unsized::<[u32]>`

There are no unsized locals in this program, so at the very least, the error message is incorrect.

[RalfJung]

We only emit that error here

https://github.com/rust-lang/rust/blob/8a9e696f43aee4b184be2e1e01fa1f6048d7aba1/compiler/rustc_const_eval/src/interpret/place.rs#L727-L729

so it very much looks like there is an unsized local in the generated MIR. Strange.

[RalfJung]

Yeah, this is the generated MIR (at opt-level 0):

fn forget_unsized(_1: T) -> () {
    debug t => _1;                       // in scope 0 at test.rs:4:34: 4:35
    let mut _0: ();                      // return place in scope 0 at test.rs:4:40: 4:40
    let mut _2: T;                       // in scope 0 at test.rs:5:24: 5:25

    bb0: {
        StorageLive(_2);                 // scope 0 at test.rs:5:24: 5:25
        _2 = move _1;                    // scope 0 at test.rs:5:24: 5:25
        _0 = const ();                   // scope 0 at test.rs:5:5: 5:26
        StorageDead(_2);                 // scope 0 at test.rs:5:25: 5:26
        return;                          // scope 0 at test.rs:6:2: 6:2
    }
}

So Miri is right, MIR building is introducing unsized locals here.

[RalfJung]

The fact that this code is accepted without the unsized_local feature gate is concerning, since this means we are hitting the broken alloca code path in codegen that is used for unsized locals:

define void @_ZN4test14forget_unsized17h6fbb3409c8556ed4E(ptr %0, i64 %1) unnamed_addr #0 {
start:
  %_2 = alloca { ptr, i64 }, align 8
  %t = alloca { ptr, i64 }, align 8
  %2 = getelementptr inbounds { ptr, i64 }, ptr %t, i32 0, i32 0
  store ptr %0, ptr %2, align 8
  %3 = getelementptr inbounds { ptr, i64 }, ptr %t, i32 0, i32 1
  store i64 %1, ptr %3, align 8
  %4 = getelementptr inbounds { ptr, i64 }, ptr %t, i32 0, i32 0
  %5 = load ptr, ptr %4, align 8, !noundef !2
  %6 = getelementptr inbounds { ptr, i64 }, ptr %t, i32 0, i32 1
  %7 = load i64, ptr %6, align 8, !noundef !2
  %8 = mul nsw i64 %7, 4
  %9 = alloca i8, i64 %8, align 16
  call void @llvm.memcpy.p0.p0.i64(ptr align 16 %9, ptr align 1 %5, i64 %8, i1 false)
  %10 = getelementptr inbounds { ptr, i64 }, ptr %_2, i32 0, i32 0
  store ptr %9, ptr %10, align 8
  %11 = getelementptr inbounds { ptr, i64 }, ptr %_2, i32 0, i32 1
  store i64 %7, ptr %11, align 8
  ret void
}

[RalfJung]

This is a rustc bug: rust-lang/rust#111175. Thanks for reporting!