nightly ICE when building flexstr in release mode about rust HOT 8 CLOSED

This should be fixed on nightly, awaiting a beta backport

from rust.

@Nilstrieb I'm going from #126889 (comment)

from rust.

Assigning @compiler-errors since he said he'd take a look at this after I figured out that this is a mess to fix.

The tl;dr here is that pre-monomorphization we wind up trying to instance drop glue.

The logic for creating drop glue should be able to assume that it is in a post-monomorphization state where using an empty ParamEnv is fine. Because we try to instance drop glue with generic types (e.g. Foo<N>) this means that we need to actually be using the ParamEnv of the body that is instancing the drop glue instead. This turns out to be quite involved to fix and not as simple as other "pass the correct param env in" bug fixes.

The... good ish? bad? news ? is that we've been getting this wrong for much longer than my PR, my PR only surfaced this due to the fact that it relies on us providing the correct ParamEnv everywhere in the compiler which apparently is not the case :-)

edit: I suspect the principled solution here is to have two codepaths (that share logic) for generating mir shims. One for instancing stuff post mono that behaves how monomorphization should be behaving and can use an empty Reveal::All env. And one for when we want to generate the mir pre-mono and it's as if we're putting the mir inline into some other body in which cases we need to be using the ParamEnv of the body we're placing the shim's mir into`.

from rust.

Bisects to #125958, cc PR author @BoxyUwU -- ah but it's likely a duplicate of #126378 and so forth.

from rust.

this is stable and doesn't use effects, so it's probably different

from rust.

I cannot reproduce this locally

edit: oh --release is required

from rust.

To avoid breaking stuff when beta gets bumped to stable we probably want to try the following:

• Figure out exactly what's causing the drop glue generation to occur in --release and see if its viable to backport a change to stop doing that to mask the bug for a cycle
• If thats not possible try to revert Const::ty removal on beta only
• If that's not possible try land a hack that's simple to reason about and beta backport it

And then when that's done try to land the principled refactoring before the next beta cutoff

from rust.

This has been beta-backported, should show up next time a beta build is released on rustup (next day or so?)

from rust.