Motivation:

Split the JMH-positive, JDK17/JIT/GC-friendly long-string rendering piece out of #776. The original PR mixed renderer, stdlib, format, compareStrings, and Scala Native changes, and Native hyperfine was not clean enough to merge as one large PR.

Key Design Decision:

Keep this PR focused on byte-rendering long strings that contain JSON escapes. This PR does not include compareStrings, char materializer, stdlib asciiSafe/substr/join, or format char-array assembly changes from #776.

Modification:

• Add CharSWAR.findFirstEscapeChar(byte[], from, to) on JVM, Scala.js, and Scala Native.
• In BaseByteRenderer, keep the existing UTF-8 byte array for long strings, locate escape bytes, bulk-copy clean chunks with System.arraycopy, and escape only the matching bytes inline.
• Precompute the exact escaped output length before writing dirty long strings so ByteBuilder does not grow repeatedly.

JDK17 / JIT / GC Notes:

• Straight byte-array loops and System.arraycopy; no reflection or internal JDK APIs.
• Reuses the existing UTF-8 byte-array allocation from master; no extra temporary arrays beyond the existing long-string encoding.
• Clean long strings stay on the same bulk-copy fast path.
• Dirty long strings avoid falling back to whole-string char escaping.
• The JDK17 API that looks tempting here, HexFormat, is intentionally not used because per-control-char formatting would be a worse JIT/GC shape than the static hex table.

Focused Target JMH + GC:

JMH ran with the project compiled at the JDK17 level using the current Mill toolchain. Command shape:

./mill --no-server bench.runJmh sjsonnet.bench.RegressionBenchmark.main -p path=... -wi 3 -i 5 -r 3s -w 2s -f 1 -prof gc

Full JMH + GC Sweep:

All 36 regression benchmark inputs were covered. The full sweep command used:

./mill --no-server bench.runJmh sjsonnet.bench.RegressionBenchmark.main -p path="$PATHS" -wi 3 -i 5 -r 2s -w 1s -f 1 -prof gc -rf json

bench.07 needs the same larger stack that bench.runRegressions normally provides, so it was rerun separately with -jvmArgsAppend -Xss100m. The full sweep is a screening run, not a claim that this renderer-only PR improves unrelated stdlib/parser cases; several unrelated rows had obvious system/JIT outliers. There were no clear time regressions by JMH error interval overlap.

Focused Rechecks:

Rows that looked suspicious in the full sweep were rerun with longer settings. The only stable allocation concern from the raw table was bench.09; a 3-fork rerun made it neutral/slightly lower. bench.06 was also rerun because the raw full sweep showed a small allocation delta.

Scala Native Hyperfine:

Native artifacts were built with ./mill --no-server 'sjsonnet.native[3.3.7].nativeLink' on both master and this branch. Each hyperfine command loops 20 CLI invocations, uses --warmup 5 --runs 60, and the table divides the reported mean/median back to per-invocation milliseconds.

Correctness Review:

• visitLongString is only used for String values when escapeUnicode = false, matching the existing ByteRenderer path.
• UTF-8 continuation bytes are always >= 0x80, so scanning the UTF-8 byte array for ", \, or < 0x20 cannot falsely match inside a non-ASCII code point.
• Control characters U+0000 through U+001F remain single-byte UTF-8 and are emitted as the same JSON escapes as the old RenderUtils.escapeByte fallback.
• Buffer safety was rechecked: every chunk copy reads elemBuilder.arr after ensureLength, so a grow cannot leave a stale array reference.
• JVM and Native CLI parity checks against master passed for long clean ASCII, long non-ASCII, quote/backslash, and control-character mixed strings.

Rejected Splits From #776:

• Format.scala char-array assembly: not JMH-positive on current master.
• length/substr/asciiSafe/join group: substr regressed, so it should not be split out as-is.
• std.join exact-capacity builder: allocation improved in one run, but no-prof JMH regressed.
• compareStrings/SWAR group: too broad and not GC-proven for a focused first split.
• JVM String.indexOf escape scan: tiny signal only, not enough for a separate PR.

Verification:

• ./mill --no-server 'sjsonnet.jvm[3.3.7].compile'
• ./mill --no-server 'sjsonnet.jvm[3.3.7].checkFormat'
• ./mill --no-server 'sjsonnet.jvm[3.3.7].test'
• ./mill --no-server 'sjsonnet.native[3.3.7].nativeLink'
• Full JMH+GC sweep over all 36 regression benchmark inputs
• Focused JMH+GC rechecks for suspicious full-sweep rows
• Native hyperfine commands above
• JVM/Native output parity checks against master for long string escape edge cases

References:

• Split from perf: comprehensive Scala Native render pipeline optimization #776
• Base: 3a9a492899420456070fb84eaa5b89f8b7dfe1bf
• Head: ed9af56139ad6a066910483104bae165cef53d16