Further investigations for performance optimizations

[jmwoliver]

Digging through the code, some spots to investigate more:

---

1. SIMD Binary Detection (High Impact, Low Effort)
File: simd.zig, parallel_walker.zig Problem (parallel_walker.zig:446-450):

for (data[0..check_len]) |byte| {
if (byte == 0) return; // Byte-by-byte for 8KB!
}
Solution: Add containsNul() using SIMD vectors to check 16-32 bytes at a time. Expected gain: 10-15%

2. Eliminate Path Allocations (High Impact, Medium Effort)
File: parallel_walker.zig Problem (line 393):

const full_path = std.fs.path.join(alloc, &.{ work.path, entry.name }) catch continue;
Every file/directory entry causes a heap allocation. Solution: Use stack-allocated path buffer with FixedBufferAllocator:

var path_buf: [std.fs.max_path_bytes]u8 = undefined;
var fba = std.heap.FixedBufferAllocator.init(&path_buf);
// Use fba.allocator() for path joins, reset between iterations
Expected gain: 10-15%

3. Gitignore Pattern Caching (High Impact, Medium Effort)
Files: gitignore.zig, parallel_walker.zig Problem (parallel_walker.zig:374-376, 235-299): loadParentGitignores() walks up the directory tree and re-parses .gitignore files for EVERY directory processed. Solution: Create a thread-safe GitignoreCache:
Cache parsed gitignore patterns per directory
Store parent inheritance chain
Mutex-protected lookup with lock-free reads for cached entries

pub const GitignoreCache = struct {
mutex: std.Thread.Mutex,
cache: std.StringHashMapUnmanaged(CachedIgnoreState),

pub fn getOrCreate(self: *GitignoreCache, dir_path: []const u8) !*CachedIgnoreState {
    // Fast path: check cache
    // Slow path: lock, load parents, parse .gitignore, cache
}

};
Expected gain: 15-20%

4. Better Initial Work Distribution (Medium Impact, Low Effort)
File: parallel_walker.zig Problem (lines 136-154): Single root path means only thread 0 gets initial work; others must steal. Solution: Pre-scan root directory to distribute subdirectories across all threads:

if (path_idx == 1 and self.num_threads > 1) {
// Open root dir, distribute subdirs to all thread deques round-robin
}
Expected gain: 5-10%

5. SIMD Case-Insensitive Search (Medium Impact, High Effort)
Files: simd.zig, matcher.zig Problem (matcher.zig:128-147): std.ascii.toLower() called per byte. Solution: SIMD case-folding using bit manipulation:

// ASCII: uppercase = lowercase ^ 0x20 (for letters only)
const case_bit: Vec = @Splat(0x20);
// Check if in A-Z range, then OR with case_bit
Expected gain: 10-20% for -i searches (not applicable to current benchmark)

Step	Optimization	Files	Est. Gain
1	SIMD binary detection	simd.zig, parallel_walker.zig	10-15%
2	Stack-allocated path buffers	parallel_walker.zig	10-15%
3	Gitignore caching	gitignore.zig, parallel_walker.zig	15-20%
4	Initial work distribution	parallel_walker.zig	5-10%
5	SIMD case-insensitive search	simd.zig, matcher.zig	10-20%

Conservative total estimate: 35-45% improvement Note: Cross-platform implementation only (no Linux-specific syscall optimizations)

File	Changes
src/simd.zig	Add containsNul(), optionally findSubstringIgnoreCase()
src/parallel_walker.zig	Path buffer optimization, gitignore cache integration, work distribution, use SIMD binary detection
src/gitignore.zig	Add GitignoreCache struct
src/matcher.zig	Integrate SIMD case-insensitive (if implementing)

---

Another investigation found similar things but a few others:

---

Gitignore Pattern Caching (Medium Impact)
Problem: loadParentGitignores() in parallel_walker.zig:235-299 walks up the directory tree for every work item. Fix:
Use a thread-safe cache keyed by directory path
Only load each .gitignore once

Optimize ** Pattern Matching (Medium Impact)
Problem: globMatch() in gitignore.zig:72-162 uses recursion for ** patterns (line 92). Fix:
Convert to iterative approach
Cache compiled pattern state

Reduce Memory Allocation (Low Impact)
Problem: _platform_memmove (2.7%) and allocation overhead visible in trace. Fix:
Pre-allocate FileBuffer with reasonable initial capacity
Reuse buffers across files

Priority	File	Changes
P0	src/parallel_walker.zig	Add per-worker output buffers
P0	src/output.zig	Add bulk flush API
P1	src/gitignore.zig	Pattern caching, iterative **
P2	src/reader.zig	Buffer reuse