tree cf1c295a938c6d15e927a0843bf4e885bc6a1c00
parent bfffe163a23e55b944a3b4d5b9ce0ded126f414a
author Takuto Ikuta <tikuta@google.com> 1771940742 +0900
committer GN LUCI <gn-scoped@luci-project-accounts.iam.gserviceaccount.com> 1771941360 -0800

Optimize HeaderChecker file I/O by parsing files only once

Instead of scheduling a separate task for each Target that includes a
given file, this change groups targets by the file they need to check.
We then pass the list of targets to `CheckFile`, which opens the file,
reads its contents, and parses the `#include` directives exactly once.

This significantly reduces:
- File I/O (open/read/close operations are reduced to 1 per file)
- CIncludeIterator allocations and parsing overhead
- Number of tasks scheduled in WorkerPool, reducing condition variable waits
  and thread synchronization overhead.

Based on local sample profiling with `gn gen --check` on no clang
modules build, total self-time spent in system locks (WorkerPool waits)
and I/O decreased substantially.

`hyperfine` benchmarking shows ~1.23x speedup and a ~43% reduction in
`System` execution time (kernel mode threads / locks / IO overhead) on
macOS:

Benchmark 1: out/gn_main gen ...
  Time (mean ± σ):      7.689 s ±  0.049 s    [User: 27.350 s, System: 38.073 s]
  Range (min … max):    7.581 s …  7.749 s    10 runs

Benchmark 2: out/gn_fast gen ...
  Time (mean ± σ):      6.265 s ±  0.047 s    [User: 27.103 s, System: 21.643 s]
  Range (min … max):    6.149 s …  6.312 s    10 runs

Summary
  out/gn_fast ... ran
    1.23 ± 0.01 times faster than out/gn_main ...

Bug: 484862257
Change-Id: If5cb8b04674fce929d972e4759021578e79e87ed
Reviewed-on: https://gn-review.googlesource.com/c/gn/+/21220
Reviewed-by: David Turner <digit@google.com>
Commit-Queue: Takuto Ikuta <tikuta@google.com>