[BUG]: tileiras SIGSEGV on aliased advanced-index load/store around reduction or matmul

cuTile Python version

1.5.0. Both examples also fail with 1.4.0.

CUDA Toolkit version

13.3 (tileiras V13.3.36)

Which installation method does this occur on?

Pip

Describe the bug

tileiras terminates with SIGSEGV for small kernels that load through
load_advanced_indexing, perform a reduction-like operation, and write back to
an aliased slice through store_advanced_indexing. A width-32 tile and a single
row are sufficient. The crash occurs during compilation, before launch.

I found two forms in separate QR kernels: one uses ct.sum, while the other uses
a 1x1 by 1x32 ct.matmul. Both reduce to tiny standalone examples and have the
same native failure signature. I grouped them because their isolated boundary is
the same aliased advanced-index load/store path. I expected each kernel to
compile, or to receive a source diagnostic if this alias pattern is unsupported.

Minimum reproducible examples

Reduction form:

import torch
import cuda.tile as ct


@ct.kernel()
def kernel(source):
    row = ct.arange(1, dtype=ct.int32)
    matrix = ct.load_advanced_indexing(
        source, (ct.Slice(0, 1), row, ct.Slice(0, 32))
    ).reshape((1, 32))
    result = ct.sum(matrix, axis=1)
    ct.store_advanced_indexing(
        source,
        (ct.Slice(0, 1), row, ct.Slice(0, 1)),
        result.reshape((1, 1, 1)),
    )


source = torch.empty((1, 1, 32), device="cuda")
ct.launch(torch.cuda.current_stream(), (1,), kernel, (source,))

Matmul form:

import torch
import cuda.tile as ct


@ct.kernel
def kernel(matrix, right):
    row = ct.arange(1, dtype=ct.int32)
    output = ct.load_advanced_indexing(
        matrix, (ct.Slice(0, 1), row, ct.Slice(0, 32))
    ).reshape((1, 32))
    left = ct.ones((1, 1), dtype=ct.float32)
    output -= ct.matmul(left, ct.load(right, (0, 0), shape=(1, 32)))
    ct.store_advanced_indexing(
        matrix,
        (ct.Slice(0, 1), row, ct.Slice(0, 32)),
        output.reshape((1, 1, 32)),
    )


matrix = torch.empty((1, 1, 32), device="cuda")
right = torch.empty((1, 32), device="cuda")
ct.launch(torch.cuda.current_stream(), (1,), kernel, (matrix, right))

Run either script in a fresh compiler cache. Crash dumps are disabled only to
avoid the separate masking error reported in #92.

run=$(mktemp -d)
CUDA_TILE_CACHE_DIR=off \
CUDA_TILE_TEMP_DIR="$run" \
CUDA_TILE_ENABLE_CRASH_DUMP=0 \
python repro.py

Relevant log output

Both forms report:

subprocess.CalledProcessError: Command '['/usr/local/cuda/bin/tileiras',
  '/tmp/.../kernel....bytecode', '-o',
  '/tmp/.../kernel....cubin', '--gpu-name', 'sm_120',
  '-O3', '--lineinfo']' died with <Signals.SIGSEGV: 11>.

cuda.tile._exception.TileCompilerExecutionError: Return code -11
Unknown location

Directly invoking tileiras on either emitted bytecode exits 139 and produces no
cubin. Each example failed in two independent cold runs with cuTile Python 1.4.0
and once more after switching the import to 1.5.0.

Environment

OS: Ubuntu 22.04.5 LTS, Linux 6.8.0-90-generic x86_64
GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition, compute capability 12.0
Driver: 580.126.09
CUDA toolkit: 13.3; nvcc 13.3.33; tileiras V13.3.36
Python: 3.13.14
PyTorch: 2.12.0+cu130 (bundled CUDA runtime 13.0)
cuTile Python: 1.5.0; also reproduced on 1.4.0
CPU: AMD EPYC 9355, 16 vCPUs

Other details

These reductions cover three original compiler crashes:

• _partial_dots_512 (ct.sum form);
• _apply_wy_rows_512 with an 8-column panel (ct.matmul form);
• _apply_wy_rows_512 with a 16-column panel (same reduced form).

The following controls compile successfully:

• width 16 in place of width 32;
• advanced-index load followed by a normal store to another tensor;
• normal load followed by an advanced-index store;
• advanced-index load and store using distinct tensors;
• an aliased width-32 advanced-index load/store with only elementwise addition.

This is why the examples retain both aliasing and a reduction or matmul, while
removing all QR-specific logic, large tensors, and multi-block launch dimensions.

Contributing Guidelines

• I agree to follow cuTile Python's contributing guidelines
• I searched the open bugs and found no duplicate for this report