[ICE]: Cannot select XtensaISD::PCREL_WRAPPER for serde Vec deserialize on xtensa-esp32s3-espidf (regression in 1.94.0.0+)

[lukasmalkmus]

[ICE]: Cannot select XtensaISD::PCREL_WRAPPER for serde Vec deserialize on xtensa-esp32s3-espidf (regression in 1.94.0.0+)

Default cargo build --release on xtensa-esp32s3-espidf fails during LLVM SelectionDAG instruction selection when deserializing a serde-internally-tagged enum that carries Vec<{ f32 }> via serde_json::from_slice. The same source compiles cleanly on 1.93.0.0 and fails on 1.94.0.0, 1.94.0.1, 1.94.0.2, and 1.95.0.0. The regression coincides with the LLVM 20 to 21 bump in the esp-rs fork.

cargo check and cargo clippy pass. Only codegen fails.

Code

Inline reproducer, four files. Build with cargo build --release.

src/lib.rs:

use serde::Deserialize;

#[derive(Deserialize)]
pub struct Inner {
    pub x: f32,
}

#[derive(Deserialize)]
#[serde(tag = "kind")]
pub enum Outer {
    B { items: Vec<Inner> },
}

#[unsafe(no_mangle)]
pub extern "C" fn parse(json: *const u8, len: usize) -> usize {
    let bytes = unsafe { core::slice::from_raw_parts(json, len) };
    match serde_json::from_slice::<Outer>(bytes) {
        Ok(Outer::B { items }) => items.len(),
        Err(_) => 0,
    }
}

Cargo.toml:

[package]
name = "xtensa-repro"
version = "0.0.0"
edition = "2024"

[lib]
crate-type = ["staticlib"]

[dependencies]
serde = { version = "1.0.228", features = ["derive"] }
serde_json = "1.0"

.cargo/config.toml:

[build]
target = "xtensa-esp32s3-espidf"

[unstable]
build-std = ["std", "panic_abort"]

[target.xtensa-esp32s3-espidf]
linker = "ldproxy"

rust-toolchain.toml:

[toolchain]
channel = "esp"

Meta

rustc 1.95.0-nightly (95e5bda86 2026-04-15) (1.95.0.0)
binary: rustc
commit-hash: 95e5bda868c960c607597bc03ed9e8f0ad26226d
commit-date: 2026-04-15
host: aarch64-apple-darwin
release: 1.95.0-nightly
LLVM version: 21.1.3

Toolchain installed via espup install --toolchain-version 1.95.0.0.

Error output

rustc-LLVM ERROR: Cannot select: 0x10991ba80: i32 = XtensaISD::PCREL_WRAPPER
  TargetConstantPool:i32<@.LCP11_13 = internal constant [2 x float] [float -1.000000e+00, float 1.000000e+00]> 0
In function: _RINvXs3_NtCsl6TiYZRKOYz_10serde_core2deINtNtCsk3RiYsTPuMT_4core6marker11PhantomDataINtNtCsekL40WErSze_5alloc3vec3VecNtCseQGsyqO7byi_12xtensa_repro5InnerEENtB6_15DeserializeSeed11deserializeINtNtNtNtCs3Fy1Qq9mMO9_5serde7private2de7content19ContentDeserializerNtNtCsjFHir6sNDPy_10serde_json5error5ErrorEEB1S_

error: could not compile `xtensa-repro` (lib)

Failing function demangled:

<core::marker::PhantomData<alloc::vec::Vec<xtensa_repro::Inner>> as serde_core::de::DeserializeSeed>
  ::deserialize::<serde::private::de::content::ContentDeserializer<serde_json::error::Error>>

The constant pool entry [2 x float] [-1.0, 1.0] is materialized inside that monomorphization.

Backtrace

RUST_BACKTRACE=full cargo build --release

This produces no additional Rust backtrace. The failure is LLVMReportFatalError exiting rustc with status 101 directly from the backend, not via a Rust panic.

Affected channels

• Current Nightly (channel = "esp")

Bisect

Toolchain	LLVM	Status
1.93.0.0	20.1.1	builds OK
1.94.0.0	21.x	fails
1.94.0.1	21.x	fails
1.94.0.2	21.x	fails
1.95.0.0	21.1.3	fails (this report)

Trigger reduction (each confirmed necessary)

• target = xtensa-esp32s3-espidf (std). The same source as a no_std staticlib on xtensa-esp32s3-none-elf builds clean.
• #[serde(tag = "...")] on the enum (forces ContentDeserializer).
• An inner struct with at least one f32 field, inside a Vec<_> inside the tagged-enum variant.
• cargo build --release. No lto, opt-level, codegen-units, or panic overrides required. The default release profile is enough.

Possibly related (open, unlabeled)

• rustc-LLVM ERROR: Cannot select: 0x11c366ee0: i32 = Constant<4096> #275: Cannot select: Constant<4096>, same target, serde 1.0.228 + piper.
• rustc-LLVM error: Unexpected node to lower with 1.94.0.1 #276: Unexpected node to lower on 1.94.0.1, slint demo.

PCREL_WRAPPER and "constant pool" haven't been filed yet in this repo or espressif/llvm-project.

Workaround

Stay on toolchain 1.93.0.0 until a fix lands.

[MabezDev]

Does this repro on the no_std target as well? It's a lot simpler for me to investigate on the no_std target.

[lukasmalkmus]

Does this repro on the no_std target as well? It's a lot simpler for me to investigate on the no_std target.

No, doesn't repro on xtensa-esp32s3-none-elf, for me.

[wolfd]

FWIW I think this is a fix espressif/llvm-project@39a5993

[rkuester]

@MabezDev you mentioned the no_std target is easier to investigate. This does reproduce on no_std (xtensa-esp32s3-none-elf), just via a different constant type than the original report.

In bitaxeorg/bitaxe-raw (an esp-hal + embassy firmware), cargo build --release fails with the same error on a [14 x i8] string constant instead of a float array:

rustc-LLVM ERROR: Cannot select: i32 = XtensaISD::PCREL_WRAPPER
  TargetConstantPool:i32<@anon... = private unnamed_addr constant
  [14 x i8] c"ADC Read Error", align 1> 0
In function: ..._usb_task_task0...poll

The string is an error message (CommandError::Message("ADC Read Error")) that gets inlined into an embassy task's poll. cargo check and cargo clippy pass; only codegen fails.

Same bisect as the original report: builds on 1.93.0.0 (LLVM 20.1.1), fails on 1.94.0.2 and 1.95.0.0 (LLVM 21.1.3).

Since the trigger is an array constant, this looks consistent with espressif/llvm-project@39a5993 ("[Xtensa] Fix ConstantPool lowering for aggregate constants"). I have not yet tested a toolchain that includes that commit; will report back if I do.