Assertion error on arm-unknown-linux-gnueabihf for maximum/minimum float

[ehuss]

I tried this code:

#![feature(float_minimum_maximum)]
fn main() {
    let x = 1.0_f64;
    let y = 2.0_f64;
    assert_eq!(x.maximum(y), y);
}

Building for target arm-unknown-linux-gnueabihf.

I expected to see this happen: Compiles

Instead, this happened: Compiler error:

rustc: /checkout/src/llvm-project/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp:973: void (anonymous namespace)::SelectionDAGLegalize::LegalizeOp(SDNode *): Assertion `(TLI.getTypeAction(*DAG.getContext(), Op.getValueType()) == TargetLowering::TypeLegal || Op.getOpcode() == ISD::TargetConstant || Op.getOpcode() == ISD::Register) && "Unexpected illegal type!"' failed.
Aborted

cc Tracking issue #91079, which has some discussion of whether or not this should be implemented using an intrinsic, which I was a little unclear about or whether this limitation is known.

Meta

rustc --version --verbose:

rustc 1.89.0-nightly (414482f6a 2025-05-13)
binary: rustc
commit-hash: 414482f6a0d4e7290f614300581a0b55442552a3
commit-date: 2025-05-13
host: x86_64-unknown-linux-gnu
release: 1.89.0-nightly
LLVM version: 20.1.4

[ehuss]

cc @Urgau or @tgross35. Not sure if either of you would know what's going on with this. If it doesn't look like there is an easy fix, I will need to disable these tests before the next beta branch/bootstrap bump (which I think is about a month from now) since the next release will start testing these via doctests.

[tgross35]

Any chance you can bisect this? I assume it was for the recent (~1-2weeks) change to use an intrinsic for these (edit: this one, bisection probably isn't needed #140792)

[tgross35]

If we need to fix the behavior, it would be a bit less churn to delete the connection between rustc's intrinsic and LLVM's (so the fallback gets used) rather than reverting the entire thing.

But we definitely should be using intrinsics for these; since they are unstable, I'm tempted to say we should disable the failing tests only on relevant platforms rather than switching back from intrinsics. At least that would give us a better chance of catching other failing platforms before stabilization, rather than changing to intrinsics and discovering more problems after stabilization.

[ehuss]

Oh, I didn't expect this to be a new behavior. Yea, bisected to #140912 which I assume was #140792.

Yea, my plan is to just disable it for that one platform. There's no big rush, I was just curious if there was a simple fix. We still have plenty of time. If there isn't anything ready in a while, I'll post the following patch.

diff --git a/library/core/src/num/f64.rs b/library/core/src/num/f64.rs
index 81ab0f14c2b..0aef4a94ed7 100644
--- a/library/core/src/num/f64.rs
+++ b/library/core/src/num/f64.rs
@@ -943,7 +943,7 @@ pub const fn min(self, other: f64) -> f64 {
     /// This returns NaN when *either* argument is NaN, as opposed to
     /// [`f64::max`] which only returns NaN when *both* arguments are NaN.
     ///
-    /// ```
+    /// ```ignore-arm-unknown-linux-gnueabihf (see https://github.com/rust-lang/rust/issues/141087)
     /// #![feature(float_minimum_maximum)]
     /// let x = 1.0_f64;
     /// let y = 2.0_f64;
@@ -970,7 +970,7 @@ pub const fn maximum(self, other: f64) -> f64 {
     /// This returns NaN when *either* argument is NaN, as opposed to
     /// [`f64::min`] which only returns NaN when *both* arguments are NaN.
     ///
-    /// ```
+    /// ```ignore-arm-unknown-linux-gnueabihf (see https://github.com/rust-lang/rust/issues/141087)
     /// #![feature(float_minimum_maximum)]
     /// let x = 1.0_f64;
     /// let y = 2.0_f64;

[tgross35]

@Urgau opened llvm/llvm-project#139380 and llvm/llvm-project#139381 so you could probably use that as a template to open an issue upstream. But it would take a while for that fix to bubble up to us, so the patch LGTM if you wind up posting it, assuming the other float types are somehow not broken on that platform.

[Urgau]

Opened llvm/llvm-project#140445 about it.