Skip to content

perf(fence): avoid repeated spec parsing and pointless subtree scans #424

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lws2004 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

perf(fence): avoid repeated spec parsing and pointless subtree scans #424

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
50 changes: 43 additions & 7 deletions src/client/dom-fence.tsx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -221,6 +221,14 @@ const SWEEP_MS = 1000
* a once-deferred scan also runs to completion to avoid starvation. This is
* not a hard frame-time cap. Remaining candidates wait for the next sweep. */
const SWEEP_BUDGET_MS = 8
/** 批量接管预算（ms）：待接管队列里堆着多条围栏，说明正处在会话切换/历史
* 加载这种「一次出现几十条」的场景。此时每轮仍只给 8ms 会让接管被帧率拖住
* （每个围栏的 React 挂载本身就要几十毫秒，一帧只够一条，而浏览器还要排队
* 绘制），30 条围栏实测要近 5 秒。这种情况下把预算放宽到 40ms，让一帧多推进
* 几条——用户此时本来就在等页面加载，短暂的主线程占用远好过几秒的空白。 */
const BULK_SWEEP_BUDGET_MS = 40
/** 队列超过这个长度就按批量场景处理。1-2 条是流式写的常态，不该放宽预算。 */
const BULK_TAKEOVER_THRESHOLD = 4
/** Rows re-checked by {@link rotationSweep}, and how many slow ticks apart
* those passes run. The observer sees structure, text, and `data-streaming`,
* so the belt-and-braces pass only has to cover changes it cannot see — in
Expand Down Expand Up @@ -432,12 +440,15 @@ function labelTextOf(block: Element): string {
* @param raw - 未修改的围栏正文。
* @returns 正文能按原有 GenUI 规范直接识别时返回 true。
*/
function isGenericGenuiFence(block: Element, raw: string): boolean {
const row = block.closest<HTMLElement>(ASSISTANT_FLOW_ROW)
if (row === null || row.dataset.chatGroupPart === 'reasoning') return false
// domLanguageOf 已反转（#258）：banner 是已知真实语言 ⇒ 不是通用块；空标签或
// 本地化通用词 ⇒ null，进入内容校验。这里不再要求标题命中任何白名单。
if (domLanguageOf(block) !== null || !block.querySelector('[data-code-block-banner]')) return false
/** 正文解析缓存：同一段围栏正文会在每一轮扫描里被反复解析校验（发现、序号
* 计算、接管各一次），一个 30 条围栏的会话每帧要做几十次 JSON.parse + 规范
* 校验 + 两次 JSON.stringify，切会话时接管时间因此随围栏数非线性恶化。
* 按正文文本缓存「这段正文是不是合规 GenUI 规范」，命中即跳过解析。
* 上限 500 条并整表清空：正文有界，会话长度无界。 */
const genericFenceBodyCache = new Map<string, boolean>()

/** 解析一段围栏正文是否符合 GenUI 规范。与前缀匹配无关，纯文本判定。 */
function parseGenericFenceBody(raw: string): boolean {
const repaired = repairFenceJson(raw)
const candidate = repaired === null ? raw : repaired.text
let value: unknown
Expand All @@ -458,6 +469,24 @@ function isGenericGenuiFence(block: Element, raw: string): boolean {
return JSON.stringify(normalizeGenuiSpec(value).value) === JSON.stringify(value)
}

function isGenericGenuiBody(raw: string): boolean {
const cached = genericFenceBodyCache.get(raw)
if (cached !== undefined) return cached
const result = parseGenericFenceBody(raw)
if (genericFenceBodyCache.size >= 500) genericFenceBodyCache.clear()
genericFenceBodyCache.set(raw, result)
return result
}

function isGenericGenuiFence(block: Element, raw: string): boolean {
const row = block.closest<HTMLElement>(ASSISTANT_FLOW_ROW)
if (row === null || row.dataset.chatGroupPart === 'reasoning') return false
// domLanguageOf 已反转（#258）：banner 是已知真实语言 ⇒ 不是通用块；空标签或
// 本地化通用词 ⇒ null，进入内容校验。这里不再要求标题命中任何白名单。
if (domLanguageOf(block) !== null || !block.querySelector('[data-code-block-banner]')) return false
return isGenericGenuiBody(raw)
}

/** 读取宿主代码块或纯文本标签中的围栏正文。 */
function rawOf(block: Element): string {
const tagTextBody = tagTextBodyOf(block)
Expand Down Expand Up @@ -1342,6 +1371,12 @@ export function installDomFenceRenderer(
}
}
if (!structural) return
// 廉价预筛：整棵子树里既没有 <pre>、也没有 tag-text 形态的 <dsh-ui> 时，
// 它不可能含有围栏。会话切换时宿主把整篇 transcript 的每个消息行都标成
// dirty，其中绝大多数是纯文本与工具卡片——对它们逐个跑 findFenceCandidates
// 是纯浪费，也是批量接管慢的主要来源。下级节点后插入 pre 时会再次触发
// mutation，所以这里跳过不会漏掉围栏。
if (element.querySelector('pre') === null && !(element.textContent ?? '').includes('<dsh-ui>')) return
// A changed node can carry a fence the node itself is not inside yet: a row
// or a wrapper that arrived whole. `includeScope` makes the scan cover the
// node and its descendants, so no shape depends on the rotation pass.
Expand Down Expand Up @@ -1712,7 +1747,8 @@ export function installDomFenceRenderer(
sourceChanged = false
requeueAwaitingSource()
}
const deadline = performance.now() + SWEEP_BUDGET_MS
const budget = pendingCandidates.size > BULK_TAKEOVER_THRESHOLD ? BULK_SWEEP_BUDGET_MS : SWEEP_BUDGET_MS
const deadline = performance.now() + budget
resolveDirtyTargets(deadline)
drainCandidates(deadline)
keepDraining()
Expand Down