[Code scan] Bound-check Tensor::inner_most_ptr for rank-1 and one-past indices

[njzjz]

This issue is a result of a Codex global repository scan.

Tensor::inner_most_ptr() is documented as supporting tensors with rank <= 2, but its bounds check has two issues:

• index == dim_size is accepted because the check uses > instead of >=, returning a one-past pointer.
• For rank-1 tensors, it evaluates shape_.dim_size(shape_.ndim() - 2) before the rank-1 branch, which underflows the dimension expression.

abacus-develop/source/source_base/module_container/ATen/core/tensor.h

Lines 397 to 408 in 84ca04b

	template <typename T>
	T* inner_most_ptr(const int &index) const {
	if (shape_.ndim() > 2) {
	throw std::invalid_argument("Invalid call, inner_most_ptr only support tensor rank <= 2!");
	}
	if (index > shape_.dim_size(static_cast<int>(shape_.ndim() - 2))) {
	throw std::invalid_argument("Invalid index, index of the inner-most must less than the inner-most shape size!");
	}
	if (shape_.ndim() == 1) {
	return data<T>() + index;
	}
	return data<T>() + index * shape_.dim_size(static_cast<int>(shape_.ndim()) - 1);

Relevant code:

if (index > shape_.dim_size(static_cast<int>(shape_.ndim() - 2))) {
    throw std::invalid_argument(...);
}
if (shape_.ndim() == 1) {
    return data<T>() + index;
}
return data<T>() + index * shape_.dim_size(static_cast<int>(shape_.ndim()) - 1);

Suggested fix:

Reject rank 0, choose the bound from shape_.dim_size(0) for rank 1 and rank 2 row access, and check index < 0 || index >= bound before returning a pointer.