R-universe r-polars build fails because rust-polars is no longer backwards compatible with rustc 1.65.0 stable

[sorhawell]

R-universe build fails
https://github.com/r-universe/rpolars/actions/runs/4884403053

I tried to downgrade to rust 1.65 stable locally, but I got no errors so far.
Something with simd is mentioned in some of the compiler errrors, but not sure what it is about.

EDIT: In a clean build I was able to reproduce the error by downgrading to rustc 1.65 stable

[sorhawell]

why R-universe fails

Upstream rust-polars has adopted crate argminmax 0.6.1 which shows to not be backwards compatible with rustc 1.65.0 stable but only 1.66 stable. The minimal previous minimal rustc version was 1.62 for polars.

argminmax 0.6.1 currently needs rustc >= 1.66.0 to not give the following compiler error

    Compiling argminmax v0.6.1
  thread 'rustc' panicked at 'assertion failed: key.param_env.is_const()', compiler/rustc_const_eval/src/const_eval/eval_queries.rs:270:5
  stack backtrace:
     0:     0x7f75faf69a43 - <std::sys_common::backtrace::_print::DisplayBacktrace as 
     .
     .
     .

possible solutions (in likely preferred ordering):

• kindly ask argminmax to become backwards compatible with 1.65 stable
• kindly ask polars to feature gate argminmax crate
• kindly ask R universe to upgrade to rustc 1.66.0 (problematic because a CRAN release would then be more time away)
• fork rust-polars where argminmax was never adopted (could work in the short run, would blow up over time)

other follow-up

• we ourselves should probably add an "R-universe-like" build to our regular PR checks to discover such backwards compatibility issues earlier.

[sorhawell]

error reproduced in hello_r_universe

[eitsupi]

Thanks for taking a look at this.

Perhaps Debian testing will skip Rust 1.64 and 1.65, so this problem will be solved when we can build this package in CRAN machine anyway.
https://tracker.debian.org/pkg/rustc

we ourselves should probably add an "R-universe-like" build to our regular PR checks to discover such backwards compatibility issues earlier.

I too thought that the handling of SIMD disabled should be tested more.

If someone wants to work on it, one way is to rewrite Cargo.toml by detecting the available Rust state by the configure script like the gifski package, I think.
https://github.com/r-rust/gifski/blob/68911d927fb34c43dede24ce1baab73346e8bf04/configure#L11-L22

[sorhawell]

Perhaps Debian testing will skip Rust 1.64 and 1.65, so this problem will be solved when we can build this package in CRAN machine anyway.

R-universe build can run on 1.65.0 when setting opp-level=1 specifically for argminmax. Probably we should cross our fingers for Debian skipping 1.65, and otherwise we can mitigate issue.

[sorhawell]

seems to work alright now

[eitsupi]

Related to #279

[eitsupi]

Debian testing migration to Rust 1.66 is in progress.
This means that if we want to do a CRAN release, MSRV is sufficient for Rust 1.66.

However, Ubuntu is currently Rust 1.65, so if we consider binary installation with the Posit Package Manager, we will need to resolve this issue and support Rust 1.65.