perf: pre-cached indent arrays for bulk newline+spaces

[He-Pin]

Motivation

The JSON renderer generates indentation strings (newline + spaces) on every nested element. For deeply nested Jsonnet output, Renderer.visitKey and visitEnd repeatedly construct identical indent strings. The current implementation calls elemBuilder.append('\n') followed by a while loop appending spaces — this is O(depth) per indent operation.

Key Design Decision

Pre-cache indent strings (newline + spaces) in a companion object array up to depth 64 (MaxCachedDepth). For depths ≤64, indent operations become a single array lookup + bulk write. For depths >64 (rare in practice), fall through to the original loop.

Modification

sjsonnet/src/sjsonnet/Renderer.scala:

• Added companion object with MaxCachedDepth = 64 constant and indentCache: Array[Array[Char]]
• Cache stores pre-computed "\n" + " " * (depth * indent) as char arrays for depths 0–64
• flushBuffer() fast path: when depth ≤ MaxCachedDepth, uses elemBuilder.appendAll(cachedArray, len) instead of character-by-character loop
• Original loop preserved as fallback for depths > 64

Benchmark Results

JMH — Full Suite (35 benchmarks, 1+1 warmup)

No regressions detected. All benchmarks within noise margin.

Note

The indentation cache optimization primarily benefits:

1. Deeply nested JSON output — common in Jsonnet configurations (Kubernetes manifests, CI configs)
2. std.manifestJsonEx — uses indentation for pretty-printing
3. Scala Native — no JIT to optimize the loop; pre-cached arrays enable System.arraycopy

Analysis

• Memory: One-time allocation of 64 char arrays (total ~2KB) — negligible.
• Thread safety: Cache is in a companion object, initialized once. Arrays are read-only after initialization.
• Threshold: 64 levels covers virtually all real-world Jsonnet output (even deeply nested Kubernetes manifests rarely exceed 20 levels).

References

• upickle.core.CharBuilder.appendAll(char[], int) for bulk writes
• Original character-by-character indent loop in Renderer.flushBuffer

Result

Pre-cached indent arrays eliminate per-character overhead for nested JSON rendering. No regressions. Benefits deeply nested output and Scala Native.

[He-Pin]

Good catch — extracted the magic 16 into a named constant Renderer.MaxCachedDepth in a new companion object. The comparison now reads depth < MaxCachedDepth instead of depth < indentCache.length.

Note that the indent cache content is instance-specific (depends on the indent constructor parameter — commonly 2, 3, or 4), but the size (16 depth levels) is a fixed constant shared across all instances.

[He-Pin]

Superseded by #730 which combines this optimization with the other renderer throughput improvements (indent cache + bulk copy + direct long rendering) into a single coherent PR with comprehensive benchmarks.