Files

T

Antigravity Agent 9a135bd19d docs(10_Wiki): Topic_Business/General/Graphic/Programming을 Topics/ 하위로 이동

최상위 10_Wiki/Topic_*였던 4개 카테고리 폴더를 10_Wiki/Topics/Topic_* 로 재배치.
콘텐츠 변경 없음(순수 폴더 이동) — Topics/ 하위 나머지 폴더는 이미 지난 커밋에서
전부 정리된 상태(잔존 항목은 에이전트 운영 상태 및 사용자가 보존을 요청한
업데이트0615/무제 3.canvas 뿐).

2026-07-05 00:44:01 +09:00

3.9 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

title

Branch Prediction

매 한 줄

"매 modern CPU 의 IPC 핵심 mechanism — 그리고 Spectre 의 attack surface.". Branch predictor 는 conditional/indirect branch 의 결과를 speculatively execute 함으로써 deep pipeline 의 stall 을 회피; 매 misprediction penalty 는 15-20+ cycles, 매 mispredicted speculative window 가 Spectre v1/v2 의 leak vector.

매 핵심

매 Predictor 종류

Static: forward-not-taken, backward-taken (hint).
Bimodal (2-bit): per-PC saturating counter.
Local history: per-branch shift register.
Global history (gshare/GAg): shared GHR XOR PC.
Tournament: meta-predictor selects local vs global.
TAGE / ITTAGE: tagged geometric history (modern SOTA).
Perceptron: AMD Zen — neural-style predictor.

매 Indirect Branches

BTB (Branch Target Buffer) — caches target.
ITTAGE for indirect.
Critical for vtables, function pointers, switch.

매 응용

Hot loops — predictable branches → near-zero penalty.
Sorted-data effect (the famous SO question).
Spectre/BranchScope side-channel attacks.
JIT branch ordering decisions (V8, Hotspot).

💻 패턴

Likely / Unlikely Hints (C++20)

if (x > 0) [[likely]] {
    fast_path();
} else [[unlikely]] {
    slow_path();
}

Branchless via Mask

// branchful
int abs_v(int x) { return x < 0 ? -x : x; }

// branchless
int abs_v_nb(int x) {
    int mask = x >> 31;          // 0 or -1
    return (x ^ mask) - mask;
}

Branchless Min via cmov

int min_v(int a, int b) {
    return a < b ? a : b;   // compilers emit cmov on x86
}

Sort Before Loop (classic)

std::sort(data.begin(), data.end());  // 정렬 → branch predictable
long sum = 0;
for (int v : data) if (v >= 128) sum += v;

LLVM `__builtin_expect`

if (__builtin_expect(error_flag, 0)) {
    handle_error();
}

Spectre v1 Mitigation (lfence)

// vulnerable
if (idx < arr_size) {
    secret = arr[idx];           // mispredict → leaks
}
// hardened
if (idx < arr_size) {
    asm volatile("lfence" ::: "memory");
    secret = arr[idx];
}

perf Branch Stats

perf stat -e branches,branch-misses ./bench
# branch-miss rate > 5% → suspect; > 10% → critical

매 결정 기준

상황	전략
Hot loop, predictable	Trust predictor — natural code
Hot loop, unpredictable	Branchless / mask / cmov
Cold path	Doesn't matter — clarity > tricks
Indirect-heavy (vtable)	Devirtualize, monomorphize
Security-sensitive	lfence / speculative load hardening

기본값: write clear branchful code; branchless only when profiler shows misprediction hotspot.

🔗 Graph

부모: Pipeline
응용: JIT Compilation
Adjacent: Spectre · Speculative Execution

🤖 LLM 활용

언제: explain mispredict cost, generate branchless equivalents, suggest hints. 언제 X: micro-arch tuning without perf data — speculation hurts.

❌ 안티패턴

Random unpredictable branch in hot loop: 20-30% slowdown.
Premature branchless conversion: hurts cold/clarity paths.
Trusting [[likely]] blindly: PGO data > human guess.
Ignoring Spectre in untrusted-input parsers: real CVE risk.

🧪 검증 / 중복

Verified (Hennessy & Patterson 6th, Agner Fog optimization manuals, Intel SDM).
신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — predictor types, branchless patterns, Spectre

3.9 KiB Raw Blame History