Adding `lto = true` causes duplicated symbol errors on `.weak` symbols

[rslawson]

Here's a minimal reproducer for the issue described in the title.

.cargo/config.toml:

[build]
target = "riscv32imc-unknown-none-elf"

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

memory.x (exact contents unimportant, but provided for convenience):

MEMORY
{
  IMEM : ORIGIN = 0x80000000, LENGTH = 1024K
  DMEM : ORIGIN = 0x10000000, LENGTH = 256K
}

REGION_ALIAS("REGION_TEXT", IMEM);
REGION_ALIAS("REGION_RODATA", DMEM);
REGION_ALIAS("REGION_DATA", DMEM);
REGION_ALIAS("REGION_BSS", DMEM);
REGION_ALIAS("REGION_HEAP", DMEM);
REGION_ALIAS("REGION_STACK", DMEM);

build.rs:

use std::env;
use std::fs;
use std::path::Path;

fn main() {
    let out_dir = env::var("OUT_DIR").expect("No out dir");
    let dest_path = Path::new(&out_dir).join("memory.x");
    fs::write(dest_path, include_bytes!("memory.x")).expect("Could not write file");

    if env::var("CARGO_CFG_TARGET_ARCH").unwrap() == "riscv32" {
        println!("cargo:rustc-link-arg=-Tmemory.x");
        println!("cargo:rustc-link-arg=-Tlink.x");
    }
    println!("cargo:rustc-link-search={out_dir}");

    println!("cargo:rerun-if-changed=memory.x");
    println!("cargo:rerun-if-changed=build.rs");
}

src/main.rs:

#![no_std]
#![no_main]

use riscv_rt::entry;

#[entry]
fn main() -> ! {
    loop {}
}

#[panic_handler]
fn panic_handler(_: &core::panic::PanicInfo) -> ! {
    loop {}
}

#[unsafe(export_name = "ExceptionHandler")]
fn exception_handler(_: &riscv_rt::TrapFrame) -> ! {
    loop {}
}

#[unsafe(export_name = "DefaultHandler")]
fn default_handler(_: &riscv_rt::TrapFrame) -> ! {
    loop {}
}

Cargo.toml:

[package]
name = "lto-bug-repro"
version = "0.1.0"
edition = "2021"

[profile.release]
lto = true # Commenting out this line should allow it to compile successfully

[dependencies]
riscv-rt = "0.13.0"

As is documented in riscv-rt, ExceptionHandler and DefaultHandler are weak symbols, and I verified in the source code for the crate that they're marked with .weak in global_asm! (relevant src). What I don't understand is why enabling lto = true causes the build to fail with the errors

error: DefaultHandler changed binding to STB_GLOBAL

error: symbol 'DefaultHandler' is already defined

error: ExceptionHandler changed binding to STB_GLOBAL

error: could not compile `lto-bug-repro` (bin "lto-bug-repro") due to 3 previous errors

I also have no idea how to track down what's causing this.

[bjorn3]

I think what is happening here is that fat LTO merges the cgu containing the global asm block and the cgu containing your definition of these symbols. During this merging LLVM is unable to remove the weak definition in the global asm block as LLVM by design doesn't touch any asm until codegen. The end result is that you get two definitions of a single symbol in the same cgu, which is not allowed even with weak symbols. Even if it was allowed, LLVM wouldn't know which of the two definitions is the weak definition. .weak ExceptionHandler defines that the symbol in the current cgu will be a weak symbol without attaching this property to either the definition inside the global asm block or the one you defined yourself.

The only way this can work I think is if the riscv-rt crate replaced the weak definition inside a global asm block with #[linkage = "weak"] on a regular function. This enables LTO to discard the weak definition when merging cgu's. It is unfortunately only possible to use this on nightly however.