rust-cuda sha2_crates_io example crashes with LLVM IR verification error: "Attribute 'align 16' applied to incompatible type" on __multi3

[stevefan1999-personal]

zluda_trace logs (tarball/zip file)

No response

Description

When attempting to run the sha2_crates_io example from rust-cuda with ZLUDA, the application crashes with a panic in the PTX compilation phase. The error indicates issues with LLVM IR generation:

• Invalid align attribute applied to __multi3 function pointer
• Terminator found in the middle of a basic block

Expected behavior: The sha2_crates_io example should compile and run successfully using ZLUDA as a CUDA replacement.

Actual behavior: The application panics during PTX module loading with LLVM verification errors, specifically:

Attribute 'align 16' applied to incompatible type!
ptr @__multi3
Terminator found in the middle of a basic block!
label "9978"

The panic originates from ptx/src/pass/llvm/emit.rs:76:13 during module load (cuModuleLoadDataEx).

Steps to reproduce

1. Clone rust-cuda repository: git clone https://github.com/Rust-GPU/rust-cuda.git
2. Navigate to the directory: cd rust-cuda
3. Build ZLUDA from source or download the latest release
4. Set LD_LIBRARY_PATH to point to ZLUDA libraries
5. Run the sha2 example:

   LD_LIBRARY_PATH="/path/to/ZLUDA/target/debug:$LD_LIBRARY_PATH" cargo run -p sha2_crates_io

6. The application crashes with the error shown above

ZLUDA version

master

Operating System

Rocky Linux 9 (with CUDA 12 and ROCm 6.4.3)

GPU

RX 6600 XT

[stevefan1999-personal]

Related #561

[vosen]

Thank you. This is now fixed in #563.
Note that even with the fixes, at least on my PC the ptx is still invalid (imports declared, but undefined function __multi3). It fails compilation on both ZLUDA and NVIDIA's CUDA.
PTX below:
sha.txt

[stevefan1999-personal]

Thank you. This is now fixed in #563. Note that even with the fixes, at least on my PC the ptx is still invalid (imports declared, but undefined function __multi3). It fails compilation on both ZLUDA and NVIDIA's CUDA. PTX below: sha.txt

Yep, problem is from here too:

ZLUDA/comgr/src/lib.rs

Lines 262 to 266 in 7a6bbc4

	let exec_data_set = comgr.do_action(
	ActionKind::CompileSourceToExecutable,
	&compile_to_exec,
	&linked_data_set,
	)?;

[stevefan1999-personal]

Looks like the LLVM compiler rt builtin is not included in

[vosen]

ZLUDA's LLVM does everything correctly and there's no issue. NVIDIA's CUDA also returns an error on cuModuleLoadDataEx.
It's rust-cuda's LLVM that fails to include __multi3