Files
2nd/10_Wiki/Topics/Architecture/Pointer_Compression.md
T
Antigravity Agent f8b21af4be Wiki cleanup: error-doc removal, dedup merge, link normalization
10_Wiki/Topics 대규모 정리:
- 오류 캡처/미완성 stub 문서 227개 제거
- 교차폴더 중복 43클러스터 병합 (63파일 → redirect)
- 링크명 정규화: 깨진 링크 수정·redirect 직결·개념 매핑 ~2,400건
- 카테고리 MOC 6개 신규 생성
- Graph 섹션 미해결 related-keyword 링크 10,058건 제거

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-20 23:52:15 +09:00

5.6 KiB
Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack
id title category status canonical_id aliases duplicate_of source_trust_level confidence_score verification_status tags raw_sources last_reinforced github_commit tech_stack
wiki-2026-0508-pointer-compression Pointer Compression 10_Wiki/Topics verified self
V8 pointer compression
compressed pointers
32-bit pointers in 64-bit heap
none A 0.9 applied
v8
javascript-engine
memory
optimization
2026-05-10 pending
language framework
cpp v8

Pointer Compression

매 한 줄

"매 64-bit 시스템 위에서 heap pointer를 32-bit 로 줄여 절반 공간으로 저장". 매 V8(2020), Hermes, JVM(CompressedOops, 2008부터) 의 heap memory ~40% 절감 + 매 cache locality 향상. 매 4GB 안 heap 제한 대신 매 typical web app 에는 충분.

매 핵심

매 동기

  • 매 64-bit OS 의 V8 heap object 의 80% 가 pointer field 차지.
  • 매 pointer = 매 8 bytes 인데 실제 정보 entropy ~32 bits (heap < 4GB).
  • 매 절반으로 → 매 RAM↓, 매 GC scan 시간 ↓, 매 L1/L2 cache hit ↑.

매 V8 scheme (post-2020)

  • Cage: 매 4GB 정렬된 가상 메모리 reservation.
  • Base register / R13: 매 cage base 주소 보관 (cached).
  • 32-bit slot: 매 cage 안 offset.
  • Smi tagging: 매 31-bit integer 의 inline (1 LSB tag).
  • Decompress: addr = base + (int64_t)(int32_t)compressed.

매 trade-off

  • + Heap mem ~3043% 절감, 매 V8 startup ↓.
  • + Cache pressure ↓ → 매 throughput 5-10% ↑.
  • Per-access decompression cost (매 작음, 보통 +1 ALU).
  • Heap 4GB 한계 (매 큰 app 의 큰 process 분할 필요).
  • External pointer 는 sandboxed pointer table 로 indirection 추가.

매 비슷한 system

  • JVM CompressedOops: 매 8-byte aligned + shift → 매 32GB heap 까지 32-bit.
  • CHERI capabilities: 매 다른 방향 — 매 capability + bounds 로 fat pointer.
  • Hermes (RN): 매 32-bit pointer + handle table.
  • WebAssembly memory64 vs memory32: 매 동일 motif.

💻 패턴

Compress / decompress (V8 conceptual)

// 매 V8 cage base = 4GB-aligned
inline Address Decompress(uint32_t compressed, Address cage_base) {
    return cage_base + static_cast<int64_t>(static_cast<int32_t>(compressed));
}

inline uint32_t Compress(Address full) {
    return static_cast<uint32_t>(full); // 매 lower 32 bits
}

// 매 Smi (Small Integer) tagging
constexpr int kSmiTagSize = 1;
inline bool IsSmi(uint32_t v) { return (v & 1) == 0; }
inline int32_t SmiValue(uint32_t v) { return static_cast<int32_t>(v) >> 1; }

Tagged slot in heap object

class HeapObject {
    uint32_t map_;           // 매 type info, compressed
    uint32_t properties_;    // 매 compressed
    uint32_t elements_;      // 매 compressed
    // 매 instance fields follow, all 4 bytes each
};
// 매 compare: 64-bit pointer system 에서 same object = 24 bytes header

JVM CompressedOops

# 매 default ON when heap ≤ 32GB on 64-bit JVM
java -XX:+UseCompressedOops -Xmx16g App
# 매 8-byte alignment + 3-bit shift → 매 32-bit reference 가 32GB 까지 cover
java -XX:ObjectAlignmentInBytes=8 ...

Sandboxed external pointers (V8 2024+)

// 매 untrusted external ptr → 매 indirection table
class ExternalPointerTable {
    Address entries_[N];
public:
    uint32_t Allocate(Address external) {
        uint32_t handle = next_++;
        entries_[handle] = external;
        return handle; // 매 store handle 32-bit, not raw ptr
    }
    Address Resolve(uint32_t handle) { return entries_[handle]; }
};

Disable pointer compression (V8 build)

# 매 large heap (>4GB) needed
gn args out.gn/x64.release
# v8_enable_pointer_compression = false
ninja -C out.gn/x64.release v8

Hermes 32-bit handle (React Native)

// 매 HermesValue = 64-bit NaN-boxed; pointer 부분은 cage offset 32-bit
struct HermesValue {
    uint64_t raw;
    bool isPointer() const { return (raw & kTagMask) == kPointerTag; }
    HeapPtr getPointer() const { return base_ + (raw & kPayloadMask); }
};

매 결정 기준

상황 Approach
Browser tab / Node ≤ 4GB V8 pointer compression ON (default)
Large server-side V8 (>4GB) Compile with compression OFF or use isolates
JVM ≤ 32GB heap CompressedOops ON (default)
JVM > 32GB UseCompressedOops auto-OFF
Mobile JS (Hermes) Always 32-bit handle
Custom GC engine 매 cage allocator + tagged 4-byte slot 직접 구현

기본값: 매 modern engine 의 default ON; 매 disable 은 명백한 large-heap 이유 있을 때만.

🔗 Graph

🤖 LLM 활용

언제: 매 V8/JVM/JS-engine 의 heap memory profiling. 매 mobile/tab memory pressure. 언제 X: 매 native-only (Rust/C++) 의 단순 64-bit pointer — 매 자체 GC 없으면 의미 적음.

안티패턴

  • >4GB V8 heap with compression ON: 매 OOM crash. 매 isolate 분할 또는 disable.
  • Misaligned cage: 매 4GB align 안 되면 decompress 산술 깨짐.
  • External raw pointer in heap: 매 sandbox bypass + crash on heap relocation. → 매 external pointer table 필수.
  • Assumption that pointer = 8 bytes in tooling: 매 V8 inspector / heap snapshot tools 가 가정 하면 corrupt read.

🧪 검증 / 중복

  • Verified (V8 blog "Pointer Compression in V8" 2020, JVM CompressedOops docs, Hermes design doc).
  • 신뢰도 A.

🕓 Changelog

날짜 변경
2026-05-08 Phase 1
2026-05-10 Manual cleanup — V8 cage scheme + CompressedOops + sandboxed external