Case Study Overview: C++ to Rust Translation

What you’ll learn: Lessons from a real-world translation of ~100K lines of C++ to ~90K lines of Rust across ~20 crates. Five key transformation patterns and the architectural decisions behind them.

• We translated a large C++ diagnostic system (~100K lines of C++) into a Rust implementation (~20 Rust crates, ~90K lines)
• This section shows the actual patterns used — not toy examples, but real production code
• The five key transformations:

Before and After Metrics

Case Study 1: Inheritance hierarchy → Enum dispatch

The C++ Pattern: Event Class Hierarchy

// C++ original: Every GPU event type is a class inheriting from GpuEventBase
class GpuEventBase {
public:
    virtual ~GpuEventBase() = default;
    virtual void Process(DiagFramework* fw) = 0;
    uint16_t m_recordId;
    uint8_t  m_sensorType;
    // ... common fields
};

class GpuPcieDegradeEvent : public GpuEventBase {
public:
    void Process(DiagFramework* fw) override;
    uint8_t m_linkSpeed;
    uint8_t m_linkWidth;
};

class GpuPcieFatalEvent : public GpuEventBase { /* ... */ };
class GpuBootEvent : public GpuEventBase { /* ... */ };
// ... 10+ event classes inheriting from GpuEventBase

// Processing requires dynamic_cast:
void ProcessEvents(std::vector<std::unique_ptr<GpuEventBase>>& events,
                   DiagFramework* fw) {
    for (auto& event : events) {
        if (auto* degrade = dynamic_cast<GpuPcieDegradeEvent*>(event.get())) {
            // handle degrade...
        } else if (auto* fatal = dynamic_cast<GpuPcieFatalEvent*>(event.get())) {
            // handle fatal...
        }
        // ... 10 more branches
    }
}

The Rust Solution: Enum Dispatch

#![allow(unused)]
fn main() {
// Example: types.rs — No inheritance, no vtable, no dynamic_cast
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub enum GpuEventKind {
    PcieDegrade,
    PcieFatal,
    PcieUncorr,
    Boot,
    BaseboardState,
    EccError,
    OverTemp,
    PowerRail,
    ErotStatus,
    Unknown,
}
}

#![allow(unused)]
fn main() {
// Example: manager.rs — Separate typed Vecs, no downcasting needed
pub struct GpuEventManager {
    sku: SkuVariant,
    degrade_events: Vec<GpuPcieDegradeEvent>,   // Concrete type, not Box<dyn>
    fatal_events: Vec<GpuPcieFatalEvent>,
    uncorr_events: Vec<GpuPcieUncorrEvent>,
    boot_events: Vec<GpuBootEvent>,
    baseboard_events: Vec<GpuBaseboardEvent>,
    ecc_events: Vec<GpuEccEvent>,
    // ... each event type gets its own Vec
}

// Accessors return typed slices — zero ambiguity
impl GpuEventManager {
    pub fn degrade_events(&self) -> &[GpuPcieDegradeEvent] {
        &self.degrade_events
    }
    pub fn fatal_events(&self) -> &[GpuPcieFatalEvent] {
        &self.fatal_events
    }
}
}

Why Not Vec<Box<dyn GpuEvent>>?

• The Wrong Approach (literal translation): Put all events in one heterogeneous collection, then downcast — this is what C++ does with vector<unique_ptr<Base>>
• The Right Approach: Separate typed Vecs eliminate all downcasting. Each consumer asks for exactly the event type it needs
• Performance: Separate Vecs give better cache locality (all degrade events are contiguous in memory)

Case Study 2: shared_ptr tree → Arena/index pattern

The C++ Pattern: Reference-Counted Tree

// C++ topology library: PcieDevice uses enable_shared_from_this 
// because parent and child nodes both need to reference each other
class PcieDevice : public std::enable_shared_from_this<PcieDevice> {
public:
    std::shared_ptr<PcieDevice> m_upstream;
    std::vector<std::shared_ptr<PcieDevice>> m_downstream;
    // ... device data
    
    void AddChild(std::shared_ptr<PcieDevice> child) {
        child->m_upstream = shared_from_this();  // Parent ↔ child cycle!
        m_downstream.push_back(child);
    }
};
// Problem: parent→child and child→parent create reference cycles
// Need weak_ptr to break cycles, but easy to forget

The Rust Solution: Arena with Index Linkage

#![allow(unused)]
fn main() {
// Example: components.rs — Flat Vec owns all devices
pub struct PcieDevice {
    pub base: PcieDeviceBase,
    pub kind: PcieDeviceKind,

    // Tree linkage via indices — no reference counting, no cycles
    pub upstream_idx: Option<usize>,      // Index into the arena Vec
    pub downstream_idxs: Vec<usize>,      // Indices into the arena Vec
}

// The "arena" is simply a Vec<PcieDevice> owned by the tree:
pub struct DeviceTree {
    devices: Vec<PcieDevice>,  // Flat ownership — one Vec owns everything
}

impl DeviceTree {
    pub fn parent(&self, device_idx: usize) -> Option<&PcieDevice> {
        self.devices[device_idx].upstream_idx
            .map(|idx| &self.devices[idx])
    }
    
    pub fn children(&self, device_idx: usize) -> Vec<&PcieDevice> {
        self.devices[device_idx].downstream_idxs
            .iter()
            .map(|&idx| &self.devices[idx])
            .collect()
    }
}
}

Key Insight

• No shared_ptr, no weak_ptr, no enable_shared_from_this
• No reference cycles possible — indices are just usize values
• Better cache performance — all devices in contiguous memory
• Simpler reasoning — one owner (the Vec), many viewers (indices)

graph LR
    subgraph "C++ shared_ptr Tree"
        A1["shared_ptr<Device>"] -->|"shared_ptr"| B1["shared_ptr<Device>"]
        B1 -->|"shared_ptr (parent)"| A1
        A1 -->|"shared_ptr"| C1["shared_ptr<Device>"]
        C1 -->|"shared_ptr (parent)"| A1
        style A1 fill:#ff6b6b,color:#000
        style B1 fill:#ffa07a,color:#000
        style C1 fill:#ffa07a,color:#000
    end

    subgraph "Rust Arena + Index"
        V["Vec<PcieDevice>"]
        V --> D0["[0] Root<br/>upstream: None<br/>down: [1,2]"]
        V --> D1["[1] Child<br/>upstream: Some(0)<br/>down: []"]
        V --> D2["[2] Child<br/>upstream: Some(0)<br/>down: []"]
        style V fill:#51cf66,color:#000
        style D0 fill:#91e5a3,color:#000
        style D1 fill:#91e5a3,color:#000
        style D2 fill:#91e5a3,color:#000
    end