Copybara import of the project:

[copybara-service]

Copybara import of the project:

--
6cbdbc5 by mohammadmseet-hue mohammadmseet@gmail.com:

Fix OOB read in blockwise quantized tensor + missing bounds checks

Bug 1: xnn_define_blockwise_quantized_tensor_value_v2 accesses dims[1]
without checking num_dims >= 2. With num_dims=1, dims[1] reads past the
caller-provided array — heap buffer overflow (confirmed with ASAN).

Bug 2: reshape_resize_bilinear_operator reads dim[0..3] without
checking num_dims >= 4. Missed by commit d45f452 which added this
check to other operators.

Bug 3: reshape_rope_operator accesses dim[num_dims-3] without checking
num_dims >= 3. With num_dims < 3, unsigned underflow in index
calculation.

--
3d7a7d6 by mohammadmseet-hue mohammadmseet@gmail.com:

Fix stack buffer overflow in static_slice and missing bounds checks in copy ops

Bug 1 — Stack buffer overflow in xnn_define_static_slice (static-slice.c:275-277)

xnn_define_static_slice() uses a local int64_t signed_offsets[XNN_MAX_TENSOR_DIMS]
(6 elements) array but iterates num_dims times without checking num_dims <= 6.
With num_dims=10, the loop writes 80 bytes into a 48-byte stack buffer — 32 bytes
of attacker-controlled stack corruption. Same pattern as Bug 1 in PR #9778.

xnn_define_static_slice_v2() has the identical issue with ends[XNN_MAX_TENSOR_DIMS].

Bug 2 — Missing bounds check in xnn_define_static_slice_v3 (static-slice.c:257-258)

memcpy(node->params.slice.begins, begins, num_dims * sizeof(int64_t)) writes
num_dims*8 bytes into begins[6] (48 bytes) without bounds checking.

Bug 3 — Missing bounds check in define_copy_node (copy.c:477)

memcpy into node->params.static_reshape.new_shape.dim without checking num_dims.
Affects xnn_define_static_reshape, xnn_define_static_expand_dims,
xnn_define_fuse_dims, and xnn_define_split_dim.

--
28e1a13 by mohammadmseet-hue mohammadmseet@gmail.com:

Fix stack buffer overflow in batch matrix multiply reshape

reshape_batch_matrix_multiply_nc() uses local arrays
batch_dims_c[XNN_MAX_TENSOR_DIMS] (6 elements) and
batch_strides_c[XNN_MAX_TENSOR_DIMS] (6 elements) but iterates
num_batch_dims times without checking num_batch_dims <= 6.

With num_batch_dims=10, writes past both 48-byte stack buffers.
Attacker-controlled values from batch_dims_a/b are written to the
overflow region — controlled stack corruption.

Reachable from all public xnn_reshape_batch_matrix_multiply_nc_*
APIs (f16, f32, pf16, pf32, qs8, qd8_f32_qc8w, qp8_f32_qc8w,
qdu8_f32_qc8w, bf16_f32).

FUTURE_COPYBARA_INTEGRATE_REVIEW=#9833 from mohammadmseet-hue:fix/bmm-num-batch-dims-check 28e1a13