Add opt-in Tier-0 no-eval compiled validation fast path

bench/validator/ab.ts

@@ -0,0 +1,110 @@
+import { safeParseCompiled } from "../../packages/tskm/src/compile.ts"
+import { safeParse } from "../../packages/tskm/src/index.ts"
+import { type BenchResult, bench } from "../lib/measure.ts"
+import { createBenchmarkFixtures, runConformance } from "./conformance.ts"
+import { gateRegressions, makeConsume } from "./sink.ts"
+
+interface ABResult {
+  readonly name: string
+  readonly interpreted: BenchResult
+  readonly compiled: BenchResult
+  readonly ratio: number
+}
+
+function run(): void {
+  const conformance = runConformance()
+  console.log(`conformance: PASS casesChecked=${conformance.casesChecked}`)
+
+  const results: ABResult[] = []
+  for (const fixture of createBenchmarkFixtures()) {
+    // One consume per fixture so its reads stay monomorphic to this fixture's output shape;
+    // a single shared consume across all fixtures goes megamorphic on JSC and distorts the
+    // ratio (see makeConsume). Both A/B halves of a fixture share the SAME instance, so the
+    // comparison is fair.
+    const consume = makeConsume()
+    // This is a shared, loaded dev machine (±10-20% per-run noise). A single A/B can swing a
+    // marginal case across 1.0x purely from a load spike landing on one half. So repeat the
+    // A/B and take the MEDIAN ratio, which is stable run-to-run; the displayed ns/op is the
+    // rep whose ratio is the median.
+    const REPS = 5
+    const reps: ABResult[] = []
+    for (let r = 0; r < REPS; r++) {
+      const opts = { warmupMs: 100, sampleCount: 20 }
+      const interpreted = bench(
+        `${fixture.name}:interpreter`,
+        () => consume(safeParse(fixture.schema, fixture.input, fixture.config)),
+        opts,
+      )
+      const compiled = bench(
+        `${fixture.name}:compiled`,
+        () => consume(safeParseCompiled(fixture.schema, fixture.input, fixture.config)),
+        opts,
+      )
+      reps.push({
+        name: fixture.name,
+        interpreted,
+        compiled,
+        ratio: interpreted.nsPerOp / compiled.nsPerOp,
+      })
+    }
+    reps.sort((a, b) => a.ratio - b.ratio)
+    results.push(reps[Math.floor(REPS / 2)] as ABResult)
+  }
+
+  console.log("\n# validator A/B (Bun/JSC)\n")
+  console.log(renderABTable(results))
+
+  const gate = gateRegressions(
+    results.map((result) => ({ name: result.name, ratio: result.ratio })),
+  )
+  console.log("\n# no-regression gate (full-materializing sink)\n")
+  console.log(gate.lines.join("\n"))
+  console.log(`\nGATE: ${gate.failed ? "FAIL" : "PASS"}`)
+  if (gate.failed) {
+    process.exitCode = 1
+  }
+}
+
+function renderABTable(results: readonly ABResult[]): string {
+  const rows = [
+    [
+      "fixture",
+      "interp ns/op",
+      "compiled ns/op",
+      "ratio",
+      "interp min/p50/p99",
+      "compiled min/p50/p99",
+    ],
+    ...results.map((result) => [
+      result.name,
+      formatNs(result.interpreted.nsPerOp),
+      formatNs(result.compiled.nsPerOp),
+      `${result.ratio.toFixed(2)}x`,
+      formatTriplet(result.interpreted),
+      formatTriplet(result.compiled),
+    ]),
+  ]
+  const header = rows[0] as string[]
+  const widths = header.map((_, col) => Math.max(...rows.map((row) => (row[col] as string).length)))
+  return rows
+    .map((row, index) => {
+      const line = row
+        .map((cell, col) => (cell as string).padEnd(widths[col] as number))
+        .join(" | ")
+      if (index === 0) {
+        return `${line}\n${widths.map((width) => "-".repeat(width)).join("-|-")}`
+      }
+      return line
+    })
+    .join("\n")
+}
+
+function formatTriplet(result: BenchResult): string {
+  return `${formatNs(result.nsPerOpMin)}/${formatNs(result.nsPerOp)}/${formatNs(result.nsPerOpP99)}`
+}
+
+function formatNs(value: number): string {
+  return value.toFixed(value >= 100 ? 1 : 2)
+}
+
+run()