perf: avoid encoding intermediate source maps

[ShMcK]

Preload rewriting generates an encoded temporary source map that combineSourcemaps immediately decodes. Use generateDecodedMap() to skip that round trip. The composer already accepts decoded maps; mapping precision (hires: 'boundary') and final encoded output stay the same.

flowchart LR
  E[Preload edits] --> B["Before: generateMap()<br/>Encode intermediate map"]
  B --> D[Decode intermediate map]
  E --> A["After: generateDecodedMap()<br/>Keep decoded mappings"]
  D --> C[Compose with original chunk map]
  A --> C
  C --> F[Encode final source map]

Loading

On a large production application with 25.5 MB of source maps, 10 measured builds per variant gave:

Median	Before	After	Reduction
Import-analysis generateBundle hook	584.12 ms	511.57 ms	12.4%
Vite invocation, including startup	4.727 s	4.707 s	0.4%

The hook improved in all 10 paired comparisons. Overall build timings overlap substantially, so this does not establish a reliable full-build speedup. All 3,463 output files, including maps, were byte-identical across 22 builds including warmups.

Review the three files in this order:

1. Runtime change: change map generation and its type assertion. No new dependency, option, or API.
2. Equivalence tests: check final maps, source content, and original positions with multiple sources and relative/POSIX/Windows paths. These preserve existing behavior and pass on the baseline too.
3. Reproducible benchmark: compare encoded/decoded composition with configurable input size.

All CI checks pass, including type checking and the Linux/macOS/Windows test matrix.

Application measurement method and local verification

Vite 8.2.0, macOS arm64, Node 22.22.2. Output: 30.8 MB JavaScript and 25.5 MB source maps across 1,389 map files (uncompressed).

One warmup build per variant, then 10 measured pairs, alternating which variant runs first. Each build uses a fresh Vite process and cleared output; dependency libraries are already built. The measured bundle task does not use Turbo caching. Both variants use the same hook instrumentation.

Hook ranges: 576.69–675.49 ms before, 490.18–533.57 ms after. Full-invocation ranges: 4.641–5.254 s before, 4.533–4.882 s after.

Verification performed locally:

• 138 focused utility/import-analysis tests pass.
• Eight additional temporary fixtures (not committed) produce identical code/maps: external, inline, hidden, disabled, relative base, runtime asset URLs, CSS-only preloads, and debug IDs. Fixtures also exercise Unicode, hashbangs, repeated dynamic imports, and shared dependencies.

Run the microbenchmark, including a larger source map

After installing repository dependencies:

node --import tsx packages/vite/scripts/benchPreloadSourcemap.ts
BENCH_LINES=100000 node --import tsx packages/vite/scripts/benchPreloadSourcemap.ts

The script checks output equality and measures map generation/composition through @jridgewell/remapping. It excludes Vite's path normalization and full build pipeline.

On upstream base 9913672, macOS arm64, Node 26.8.1: five fresh processes per input size, each with five warmup pairs and 20 measured pairs in alternating order. Results below are medians of process medians; map sizes include source content.

Source	Original map	Encoded	Decoded
10,000 lines / 0.308 MB	1.82 MB	19.65 ms	18.23 ms
100,000 lines / 3.28 MB	19.37 MB	251.14 ms	198.56 ms

The larger case improved in four of five processes; one regressed (206.40 → 230.05 ms). These synthetic results are not full-build speedups or a guarantee that gains scale with map size.

AI assistance: Codex prepared the patch, tests, benchmark, and description; inspected the composition path; and ran the local checks above. Human author review is pending.