r? @Mark-Simulacrum

rustbot has assigned @Mark-Simulacrum.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

@rustbot ready

r? me

Hi! Do you have any tests or a reproducer which was broken and which this change is fixing?

Hey @ZuseZ4, thanks for reviewing!

Yeah, the ICE triggers specifically when combining the autodiff feature with Link-Time Optimization (-C lto=fat).

Here’s a minimal reproducer:

  #![feature(autodiff)]
  use std::autodiff::*;
  
  #[autodiff_reverse(df, Duplicated, Duplicated)]
  fn f(x: &[f64; 2], y: &mut f64) {
      *y = x[0] * x[0] + x[1] * x[0];
  }
  
  fn main() {
      let x = [2.0, 2.0];
      let mut bx = [0.0, 0.0]; 
      let mut y = 0.0;
      let mut by = 1.0; 
  
      df(&x, &mut bx, &mut y, &mut by);
      assert_eq!(bx, [6.0, 2.0]);
  }

If you compile that with rustc --edition 2021 -Z autodiff=Enable -C lto=fat test_autodiff.rs on the codebase before this PR, you get this Enzyme backend panic:

error: <unknown>:0:0: in function preprocess__RNvCsim5ttFnN1cV_13test_autodiff2df void (ptr, ptr, ptr, ptr): Enzyme: cannot compute with global variable that doesn't have marked shadow global
@enzyme_dup = external global ptr

Basically, because those dummy metadata globals were only declared in the C++ backend, LTO was aggressively stripping their initializers out as dead code. Moving them directly into core natively protects them from LTO, so Enzyme can actually read them.

Let me know if you want me to drop this reproducer into tests/ui or tests/run-make as an official test case, happy to push a commit for that!