LLVM Fatal Error: Copy one register into another with a different width

[AmesingFlank]

repro:

import taichi as ti

ti.init(ti.cuda)

@ti.dataclass
class S:
    i: ti.i16 = 0
    b: bool = False

sf = S.field(shape=(10,  2))

@ti.kernel
def k() -> S:
    ss = [sf[0,0], sf[0,1]]
    return ss[0]

k()

this throws:

Traceback (most recent call last):
  File "C:\Users\ldfra\Code\taichi\test_llvm_throw.py", line 19, in <module>
    k()
  File "c:\users\ldfra\code\taichi\python\taichi\lang\kernel_impl.py", line 936, in wrapped
    return primal(*args, **kwargs)
  File "c:\users\ldfra\code\taichi\python\taichi\lang\kernel_impl.py", line 868, in __call__
    return self.launch_kernel(kernel_cpp, *args)
  File "c:\users\ldfra\code\taichi\python\taichi\lang\kernel_impl.py", line 801, in launch_kernel
    raise e from None
  File "c:\users\ldfra\code\taichi\python\taichi\lang\kernel_impl.py", line 798, in launch_kernel
    prog.launch_kernel(compiled_kernel_data, launch_ctx)
RuntimeError: [llvm_context.cpp:taichi::lang::TaichiLLVMContext::{ctor}::<lambda_ea41ae353cd42ca584262925730f46b7>::operator ()@80] LLVM Fatal Error: Copy one register into another with a different width

The problem looks to be the b:bool = False member, because removing it or changing its type removes the error.

[ailzhang]

cc: @listerily , maybe this is related to the boolean type?

[listerily]

cc: @listerily , maybe this is related to the boolean type?

Seems yes. Maybe there is some special code generation rules related to field? I would look into it later.

[listerily]

This issue is caused by TaskCodeGenCUDA::create_intrinsic_load. I'll keeping on working to solve that.

[listerily]

Hi! Your issue has been solved and it has been merged into master. Additionally, default values are not supported in taichi dataclasses. Please remove default values in S like this:

@ti.dataclass
class S:
    i: ti.i16
    b: bool