`reshape(view(...))` inside a GPU kernel triggers `InvalidIRError`

[omlins]

Description

When calling reshape(view(...)) inside an AMDGPU kernel, it fails with an InvalidIRError.

Minimal Working Example (MWE)

using AMDGPU

function kernel!(A)
    i = workitemIdx().x
    B = reshape(view(A, :, i), (size(A, 1),))  # ❌ This triggers InvalidIRError
    B[1] = 9.0
    return
end

# Create a ROCArray on the host (it will be converted automatically)
A = ROCArray{Float64}(undef, 4, 4)

# Launch kernel
@roc groupsize=4 kernel!(A)
AMDGPU.synchronize()

Expected Behavior

The kernel should execute without errors, and reshape(view(...)) should produce a valid ROCDeviceArray.

Actual Behavior

The following error occurs:

ERROR: InvalidIRError: compiling MethodInstance for kernel!(::ROCDeviceArray{Float64, 2}) resulted in invalid LLVM IR
Reason: unsupported call to an external C function

System Information

• Julia Version: 1.11.2
• AMDGPU.jl Version: master

[omlins]

@pxl-th the equivalent works in CUDA.jl

[pxl-th]

@omlins I'll take a look

[omlins]

@omlins I'll take a look

Thanks!

[omlins]

@pxl-th Could you please tag a new release soon?

[pxl-th]

Yeah, I'll tag right after #740