Problem

In non-ALTREP mode (immediate materialization), vroom extracts fields one at a time:

for (size_t row = 0; row < num_rows; ++row) {
    string val = idx.get(row, col);  // Individual call per row
    result[row] = Rf_mkCharLen(val.data(), val.size());
}

This creates function call overhead and prevents optimization opportunities.

Proposed Solution

Add a bulk extraction API for entire columns:

// Instead of N individual get() calls:
std::vector<std::string_view> extract_column(size_t col) const {
    std::vector<std::string_view> result;
    result.reserve(num_rows());
    
    // Single pass through column data
    for (size_t row = 0; row < num_rows(); ++row) {
        auto span = idx_.get_field(row, col);
        result.emplace_back(buffer_ + span.start, span.length());
    }
    return result;
}

Benefits

• Pre-allocation of output vector
• Sequential memory access pattern
• Potential SIMD acceleration for field extraction
• Reduced function call overhead

Implementation Notes

• Only beneficial for non-ALTREP mode where entire columns are materialized upfront
• Should be opt-in via a flag or detected automatically based on access pattern
• Works well with the flat index design (Implement flat index array for O(1) field access libvroom#590)

Related

• Depends on libvroom flat index implementation (Implement flat index array for O(1) field access libvroom#590)
• Complements direct type parsing optimization