/profiling

Memory Profiling

Memory Footprint: What Actually Counts (WWDC21 10180)

Three categories of memory — only two count against you:

• The equation: footprint = dirty + compressed. Clean memory is free (WWDC21 10180).
• iOS has no swap — swap is macOS-specific. In vmmap output on iOS, the "swapped" column means "compressed"; only the "dirty size" and "swapped size" columns matter for footprint (WWDC21 10180).
• All malloc allocations are ≥ 16 bytes and 16-byte aligned — a 4-byte request costs 16 bytes (WWDC24 10173).
• A page is a fixed-size, indivisible chunk (16 KB OS pages) — writing to any part dirties the whole page and your process is charged for all of it (WWDC21 10180, WWDC24 10173).
• Under pressure the system compresses/swaps dirty pages, evicts clean ones, kills background tasks, then your app (WWDC24 10173).
• Why a smaller footprint matters: less likely to be terminated in background → state preserved → faster activations; more headroom for features; equal behavior on older devices with less RAM (WWDC21 10180).

Fragmentation: The ≤25% Rule (WWDC21 10180)

Fragmentation = dirty pages not 100% utilized. It's a footprint multiplier — "50% fragmentation doubled my footprint from 2 to 4 dirty pages."

• Rule of thumb: aim for about 25% fragmentation or less.
• Fixes: allocate objects with similar lifetimes close together in memory; use autorelease pools (everything allocated inside is released when the pool exits scope → similar lifetimes). Transient spikes that get freed still leave fragmentation holes (WWDC24 10173).
• Long-running processes (e.g. extensions) are especially fragmentation-prone.
• Measure: vmmap -summary, bottom section by malloc zone — the % FRAG column is % wasted per zone. Normally only DefaultMallocZone matters, but with Malloc Stack Logging enabled all heap allocations land in MallocStackLoggingLiteZone — read that one instead.

Instruments Setup

Choosing the Right Tool (WWDC24 10173)

Enable MallocStackLogging first (Scheme ▸ Diagnostics) — it records the backtrace + timestamp of every allocation, powers the memory graph's backtraces and malloc_history, and gives type context where C/C++ has none (WWDC24 10173). A device is needed for timing work; Simulator is acceptable for heap-shape analysis.

Allocations Instrument

1. Xcode > Product > Profile (Cmd+I)
2. Choose Allocations template
3. Record, reproduce the issue, stop

Leaks Instrument

• Included in the Allocations template by default
• Runs periodic scans for unreachable memory
• Red crosses in the Leaks lane indicate detected leaks

Allocations Instrument

Key Metrics

Reading the Results

• Sort by Live Bytes to find largest memory consumers
• Sort by # Living to find most numerous allocations
• Click a category to see individual allocations with stack traces
• Persistent column shows objects that survive across generations

Generation Analysis (Mark Generation)

Track memory growth over repeated operations:

1. Start recording
2. Click Mark Generation to baseline
3. Perform an operation (open screen, load data, etc.)
4. Click Mark Generation again
5. Repeat the same operation
6. Compare generations — growth between identical operations = leak or unbounded cache

What to look for:

• Stable: Each generation has similar Live Bytes
• Leaking: Each generation grows — objects from previous generations persist
• Caching: Growth that plateaus after a few generations (may be intentional)

Drilling into persistent growth (WWDC24 10173): each generation = allocations made after that mark and still alive at the end. Expand a growing generation, sort by size, drill to an individual address — then hand that address to the Memory Graph Debugger (filter by address), walk the reference chain to the owner, and use the Inspector backtrace to reach source. The session's bug: a cache key built from Date.now instead of the file's creation date meant every lookup missed, so every thumbnail was re-created and retained forever in a static cache. Doctrine: a cache that never hits is a leak with extra steps — verify key stability.

Transient Spikes (WWDC24 10173)

Spikes that get freed still hurt: memory pressure at the peak, plus long-term fragmentation holes.

• One spike: drag-select the track from low point to peak → statistics show allocations Created & Still Living in that window → sort by total bytes.
• Repeated spikes: select a wide range → lifespan filter "Created & Destroyed" → switch to the call tree to see which code creates the temporaries. The session demo found 8 GB of temporaries across three gallery opens, dominated by autorelease-pool content pages.
• Calling ObjC frameworks from Swift yields autoreleased objects; in a loop they pile into the thread's top-level pool (visible as "content pages" in Allocations) and free only when it drains — a memory cliff. Fix: nested autoreleasepool per iteration (see Autorelease Pool Optimization below).

Common Memory Issues

Unbounded Cache Growth

// Bad — cache grows forever
class ImageCache {
    var cache: [URL: UIImage] = [:]

    func image(for url: URL) -> UIImage? {
        return cache[url]
    }
}

// Good — use NSCache for automatic eviction
class ImageCache {
    private let cache = NSCache<NSURL, UIImage>()

    init() {
        cache.countLimit = 100
        cache.totalCostLimit = 50 * 1024 * 1024 // 50 MB
    }

    func image(for url: URL) -> UIImage? {
        return cache.object(forKey: url as NSURL)
    }

    func store(_ image: UIImage, for url: URL) {
        let cost = image.cgImage.map { $0.bytesPerRow * $0.height } ?? 0
        cache.setObject(image, forKey: url as NSURL, cost: cost)
    }
}

Large Images Not Downsampled

// Bad — full resolution image loaded into memory
let image = UIImage(contentsOfFile: photoPath)
// 4032x3024 photo = ~48 MB in memory

// Good — downsample to display size
func downsample(imageAt url: URL, to pointSize: CGSize, scale: CGFloat) -> UIImage? {
    let maxDimensionInPixels = max(pointSize.width, pointSize.height) * scale
    let options: [CFString: Any] = [
        kCGImageSourceCreateThumbnailFromImageAlways: true,
        kCGImageSourceShouldCacheImmediately: true,
        kCGImageSourceCreateThumbnailWithTransform: true,
        kCGImageSourceThumbnailMaxPixelSize: maxDimensionInPixels
    ]

    guard let source = CGImageSourceCreateWithURL(url as CFURL, nil),
          let cgImage = CGImageSourceCreateThumbnailAtIndex(source, 0, options as CFDictionary)
    else { return nil }

    return UIImage(cgImage: cgImage)
}

Leaks Instrument

How Leaks Detection Works

• Scans heap at intervals for objects with no root references
• Detects unreachable memory — allocated but no path from stack/globals
• Does NOT detect logical leaks (reachable but never used, like unbounded caches)

The three reachability states (WWDC24 10173): useful (reachable, will be used), abandoned (reachable but never used again — over-caching, fat singletons; wastes footprint and no tool flags it), and leaked (unreachable — cycles, lost pointers). Leaks-the-tool only finds the third kind. "Strive to have zero leaks in your app" (WWDC21 10180).

Leak-tool limits (WWDC24 10173):

• Tools classify references as strong / weak / unowned / unmanaged / conservative — raw bytes that merely look like pointers (all C, and C++ without virtual methods). Conservative references cause false negatives and non-determinism (4 leaks one run, 5 the next). Trick: run the suspect code in a loop ~100× — multiplied leaks are easy to spot.
• noreturn gotcha: locals in a function ending in dispatchMain() (noreturn) get their cleanup optimized away and are falsely reported leaked — store app-lifetime objects in a static/global instead.
• If weak/unowned edges are missing from the memory graph, set build setting Reflection Metadata Level = All.

Common Retain Cycle Patterns

Closure Capturing Self

// Bad — retain cycle: self -> timer -> closure -> self
class ViewModel {
    var timer: Timer?

    func startPolling() {
        timer = Timer.scheduledTimer(withTimeInterval: 5, repeats: true) { _ in
            self.refresh() // Strong capture of self
        }
    }
}

// Good — weak capture
timer = Timer.scheduledTimer(withTimeInterval: 5, repeats: true) { [weak self] _ in
    self?.refresh()
}

Delegate Retain Cycle

// Bad — strong delegate creates cycle
protocol ServiceDelegate: AnyObject {
    func didComplete()
}

class Service {
    var delegate: ServiceDelegate? // Strong reference
}

class ViewModel: ServiceDelegate {
    let service = Service() // ViewModel -> Service -> ViewModel
    init() { service.delegate = self }
}

// Good — weak delegate
class Service {
    weak var delegate: ServiceDelegate?
}

Closure in Collection

// Bad — closures in array capture self strongly
class Coordinator {
    var handlers: [() -> Void] = []

    func register() {
        handlers.append {
            self.handleEvent() // Each closure retains self
        }
    }
}

// Good — weak capture in each closure
handlers.append { [weak self] in
    self?.handleEvent()
}

NotificationCenter (Pre-iOS 9 / Manual)

// Modern iOS/macOS: system auto-removes observers on dealloc
// But if using block-based API, the token retains the closure:

// Bad — token keeps closure alive
let token = NotificationCenter.default.addObserver(
    forName: .dataChanged, object: nil, queue: .main
) { _ in
    self.reload() // Retained by token
}

// Good — store and remove, or use weak self
let token = NotificationCenter.default.addObserver(
    forName: .dataChanged, object: nil, queue: .main
) { [weak self] _ in
    self?.reload()
}

Closure Contexts and the Implicit-self Trap (WWDC24 10173)

• Every capturing closure allocates a closure context on the heap (1:1 with the closure); captures are strong by default. The memory graph shows contexts with references labeled only "capture" (no variable names) — double-click one to jump to the closure's source.
• Implicit self trap: assigning a method as a closure value (generator = defaultAction) strongly captures self → self-cycle. Replace with an explicit closure and a capture list.

weak vs unowned — Real Costs (WWDC24 10173)

• weak: the first weak reference to an object allocates a separate weak-reference storage side allocation, and accesses go through swift_weakLoadStrong(). At scale it matters — with 1,000,000 small objects the weak-storage overhead was about the size of the objects themselves. Still the safe default when lifetime isn't guaranteed.
• unowned: no side allocation, faster access, non-optional — but accessing after deallocation is a deterministic crash. Use only when the reference provably cannot outlive its destination (e.g. a closure stored by the very object it captures, never escaping): generator = { [unowned self] in self.defaultAction($0) }.
• Reducing reference-counting overhead generally: enable Whole-Module Optimization, specialize hot generics, and keep your most-copied structs trivial — avoid reference-type, copy-on-write (String/Array/Dictionary), and any-existential fields in hot structs. ❌ Never circumvent ARC with Unmanaged/manual retain-release to "save" refcount traffic — the resulting leaks cost far more than the overhead (WWDC24 10173).

Xcode Memory Graph Debugger

When to Use

• Leaks instrument shows leaks but you need to see why objects are retained
• You suspect retain cycles but can't find them in code review
• You need to see the full ownership graph of an object

How to Use

1. Run app in Debug mode (not Profile)
2. Reproduce the state where objects should be deallocated
3. Click the Memory Graph button in Xcode's debug bar (looks like interconnected nodes)
4. Xcode pauses and captures the heap

Reading the Graph

• Left sidebar: List of all live objects grouped by type
• Purple exclamation marks: Xcode-detected leaks / retain cycles
• Center canvas: Visual graph of object references
• Click an object to see its retain graph — who holds a strong reference to it

Debugging Tips

• Filter by your module name to ignore system objects
• Look for cycles — two objects pointing at each other
• Check backtrace in the right sidebar to see where the object was allocated
• Export the graph: File > Export Memory Graph for sharing

Enable Malloc Stack Logging

For full allocation backtraces in Memory Graph Debugger:

1. Scheme > Run > Diagnostics
2. Enable Malloc Stack Logging (Live Allocations Only)
3. Now the right sidebar shows the exact call stack that allocated each object

MSL also powers malloc_history and leaks' per-leak allocation stacks, and gives type context (e.g. "malloc in PalmTree::growCoconut()") where C/C++ allocations otherwise have none (WWDC24 10173). It has continuous per-allocation cost — enable it for investigations, not day-to-day runs.

Command-Line memgraph Triage (WWDC21 10180)

A memgraph is a snapshot of your process's address space: address + size of every VM region and malloc block, plus the pointers between them. Capture one from Xcode's memory graph debugger (File > Export Memory Graph) or get them from performance tests (below). Apple's exact triage order:

1. Always check leaks first: leaks post.memgraph. A ROOT CYCLE header = retain cycle; with MSL you also get each leak's allocation call stack.
2. No leaks? Check the heap:
   • vmmap -summary pre.memgraph vs post.memgraph → confirm the footprint delta and find which region holds it — heap objects live in regions starting MALLOC_
   • heap -diffFrom pre.memgraph post.memgraph → objects in post but not pre, by class with counts + byte sums ("non-object" = raw malloc'd bytes in Swift)
   • heap -addresses (filterable, e.g. only non-objects ≥ 500 KB) → get culprit addresses
   • leaks --traceTree <addr> → tree of objects referencing that address (best when MSL is NOT available)
   • leaks --referenceTree (+ --groupByType) → top-down reference tree with root guesses — best when you know there's a big regression but not which objects
   • malloc_history -fullStacks <addr> → the allocation call stack for that address (requires MSL)
3. Fragmentation: vmmap -summary, per-zone % FRAG column and dirty+swap frag size (see the ≤25% rule above). In Instruments, the Allocations list's "persisted" objects keep pages dirty and "destroyed" objects created the free slots — investigate both.

Memory Performance Tests: XCTMemoryMetric (WWDC21 10180)

Gate memory regressions in CI:

func testSaveMeal() {
    let app = XCUIApplication()
    let options = XCTMeasureOptions()
    options.invocationOptions = [.manuallyStart]
    measure(metrics: [XCTMemoryMetric(application: app)], options: options) {
        app.launch()
        startMeasuring()
        app.cells.firstMatch.buttons["Save meal"].firstMatch.tap()
        let savedButton = app.cells.firstMatch.buttons["Saved"].firstMatch
        XCTAssertTrue(savedButton.waitForExistence(timeout: 30))
    }
}

• Results average five iterations; set that average as a baseline so a future run failing it is a regression — "stop, investigate, and fix" (WWDC21 10180).
• Run via xcodebuild with the enablePerformanceTestsDiagnostics flag to collect ktrace files (for non-memory metrics — open in Instruments) and memgraphs (for memory metrics) attached to the .xcresult bundle.
• XCTest automatically enables malloc stack logging, and you get two memgraphs per run — pre-prefixed (start of the measured iteration) and post-prefixed (end) — ready for the heap -diffFrom pre.memgraph post.memgraph workflow above.

Autorelease Pool Optimization

In tight loops creating Objective-C bridged objects:

// Bad — autoreleased objects accumulate until loop ends
for item in largeDataSet {
    let string = item.name as NSString       // Autoreleased
    let data = string.data(using: .utf8)     // Autoreleased
    process(data)
}
// Memory spikes until loop completes and pool drains

// Good — drain pool each iteration
for item in largeDataSet {
    autoreleasepool {
        let string = item.name as NSString
        let data = string.data(using: .utf8)
        process(data)
    }
}
// Memory stays flat — pool drains each iteration

When autoreleasepool Matters

• Processing large collections (1000+ items)
• Creating temporary NSString, NSData, NSNumber objects in loops
• Image processing pipelines
• Not needed for pure Swift value types (String, Data, Array)

Memory Footprint Tracking in Code

import os

func logMemoryFootprint() {
    var info = mach_task_basic_info()
    var count = mach_msg_type_number_t(
        MemoryLayout<mach_task_basic_info>.size / MemoryLayout<natural_t>.size
    )
    let result = withUnsafeMutablePointer(to: &info) {
        $0.withMemoryRebound(to: integer_t.self, capacity: Int(count)) {
            task_info(mach_task_self_, task_flavor_t(MACH_TASK_BASIC_INFO), $0, &count)
        }
    }
    if result == KERN_SUCCESS {
        let usedMB = Double(info.resident_size) / 1_048_576
        os_log("Memory footprint: %.1f MB", usedMB)
    }
}

Why Your App Gets Killed (WWDC20 10078)

The six termination reasons — rank which actually hits your users with MXAppExitMetric (daily counts per exit cause, foreground + background) before debugging locally:

Rules that follow:

• Going to background, aim for < 50 MB footprint — the smaller, the better. Drop caches and anything re-readable from disk in didEnterBackground (WWDC20 10078).
• ❌ Don't assume a kill will reproduce under the debugger — watchdog and jetsam kills specifically never do.
• Save state on going background (text field content, media playback position — UIKit State Restoration does the heavy lifting) so a jetsam relaunch is invisible: "many users won't even realize the app had been killed."
• For sustained background CPU work, use BGProcessingTask — several minutes of runtime while charging, without CPU resource limits.

Background Task Hygiene (WWDC20 10078)

// ❌ Leaks the assertion: no name, no expiration handler, ID clobbered on reuse
self.taskID = UIApplication.shared.beginBackgroundTask()
upload(data) { self.taskID = .invalid }   // handler may never run

// ✅ Named task, local identifier, expiration handler that only cleans up + ends
var taskID: UIBackgroundTaskIdentifier = .invalid
taskID = UIApplication.shared.beginBackgroundTask(withName: "UploadPhotos") {
    // cleanup only — never kick off new work here
    UIApplication.shared.endBackgroundTask(taskID)
    taskID = .invalid
}
upload(data) {
    UIApplication.shared.endBackgroundTask(taskID)
    taskID = .invalid
}

• Always use the named variant so an unended task is identifiable in logs; always provide an expiration handler (it's safe to call endBackgroundTask inside it); call endBackgroundTask() immediately when work finishes — holding the assertion keeps the device awake.
• Budget before starting: estimate the work's duration (use 5s as the lower bound) and compare against UIApplication.shared.backgroundTimeRemaining; too little time → enqueue a BGProcessingTask instead.
• iOS 13.4+ prints a console message when a background task is held too long — watch for it during development. In the field, pair os_signpost begin/end events around the handler and compare counts in MetricKit payloads — imbalanced counts mean the handler hung or crashed.

Quick Diagnosis Checklist

• Run Allocations with Generation Analysis — does memory grow across identical operations?
• Run Leaks instrument — any detected leaks? (Run suspect code ~100× in a loop to defeat conservative-reference false negatives)
• Check Memory Graph Debugger for retain cycles (purple warnings)
• Search for self in closures without [weak self] or [unowned self]
• Verify delegates are declared weak
• Check NSCache usage instead of plain dictionaries for caches — and verify cache keys are stable (a cache that never hits is a leak with extra steps)
• Verify images are downsampled to display size
• Use autoreleasepool in tight loops with ObjC-bridged objects
• Footprint = dirty + compressed — check vmmap -summary dirty/swapped columns, not virtual size
• Fragmentation ≤ 25% per malloc zone (% FRAG in vmmap -summary)
• Background footprint < 50 MB; total < 200 MB if supporting devices older than iPhone 6s
• Every beginBackgroundTask is named, has an expiration handler, and is ended on all paths