Summary

• Pre-compile glob exclusion patterns once in globwalk_internal instead of cloning Vec<Glob> per include pattern and re-compiling each time. Adds FileIterator::not_any to wax to accept a pre-compiled FilterAny directly.
• Cache the git-root-to-package path prefix in hash_objects so each file does a cheap strip_prefix instead of full relative_path_between path component iteration.

Why

CPU profiling showed get_package_file_hashes as the dominant hotspot (538% self time). Within that path, globwalk_internal was cloning and re-compiling exclusion Glob patterns (each containing a compiled Regex) for every include glob — O(N includes × M excludes) regex clones + N recompilations per globwalk call. In hash_objects, the path prefix computation was doing redundant component iteration for every file when the relationship between git root and package path is constant.

Benchmarks

Measured with hyperfine --warmup 5 across three monorepos of different sizes:

~300 packages:

Benchmark 1 (patched):  6.584s ± 0.144s
Benchmark 2 (baseline): 7.087s ± 0.141s
  → 1.08 ± 0.03 times faster

~100 packages:

Benchmark 1 (patched):  1.378s ± 0.104s
Benchmark 2 (baseline): 1.414s ± 0.059s
  → 1.03 ± 0.09 times faster

~6 packages:

Benchmark 1 (patched):  643.8ms ± 65.1ms
Benchmark 2 (baseline): 671.1ms ± 99.8ms
  → 1.04 ± 0.19 (within noise)

Improvements scale with package count as expected since both optimizations eliminate per-package redundant work.

Testing

All existing tests pass across wax, globwalk, turborepo-scm, and turborepo-task-hash.