The Silver Searcher — ag (C)

The tool that made "grep that respects .gitignore" a category. Its algorithms are the pre-SIMD state of the art, which makes it a useful control: everything ripgrep does faster, ag does the textbook way.

Field	Value
Language	C
License	Apache-2.0
Repository	ggreer/the_silver_searcher
Surveyed revision	a61f1780b64266587e7bc30f0f5f71c6cca97c0f (2.2.0-60)
Category	Unindexed scanner
Engine class	Boyer-Moore / hashed substring for literals; PCRE for regex
Index	None
Interactive	None

Last reviewed: August 28, 2026.

---

Overview

What it solves

Recursive search that skips what version control already says to skip. ag established the default that ripgrep inherited, and its literal path is a clean two-algorithm implementation with an explicit platform gate.

How it works

The ten dimensions

1. Pattern language

PCRE for regex; a literal fast path when the pattern has no metacharacters. Case sensitivity is a tri-state (CASE_SENSITIVE, CASE_INSENSITIVE, smart).

2. Engine architecture

Two literal searchers, selected by platform capability:

/* hash_strnstr only for little-endian platforms that allow unaligned access */

On those platforms hash_strnstr runs — a rolling-hash substring search over a precomputed table; elsewhere, boyer_moore_strnstr with alpha_skip_lookup and find_skip_lookup skip tables. [source-verified] Regex falls through to PCRE per line.

This is the pre-memmem design: skip tables rather than vectorised candidate-finding, and the field's measured answer since has been that rare-byte SIMD scanning beats classical skipping on realistic text.

3. Prefilter and literal extraction

None from a regex. The literal path is chosen only when the whole pattern is literal.

4. Corpus access

mmap by default on POSIX (MapViewOfFile on Win32), with binary files skipped before mapping when binary search is disabled — the comment notes that "if not using mmap, binary files have already been skipped", so the two paths detect at different points. [source-verified]

5. Concurrency

A worker pool over files, with uthash for the ignore structures.

6. Index

None.

7–8. Result model and Unicode

Matching lines in walk order; no ranking. Unicode is whatever PCRE was built with; the literal path is byte-oriented with ASCII case folding.

9–10. Interactive behaviour and evidence

None, and none quoted. ag's historical benchmark claims predate ripgrep and are [literature].

Strengths

• Established the ignore-aware default the whole category now assumes.
• Two literal algorithms with an explicit portability gate — unaligned access and endianness are checked rather than assumed.
• Small and readable: the whole search core is a few hundred lines.

Weaknesses

• Classical skip tables rather than vectorised search, which is where the measured gap against ripgrep comes from.
• mmap by default, with the truncation hazard GNU grep withdrew and ripgrep made unsafe.
• PCRE per line for regex, with no literal extraction to avoid it.
• Maintenance has largely stopped, so it is a snapshot of 2014-era practice.

Key design decisions and trade-offs

Decision	Rationale	Trade-off
Respect .gitignore by default	Source trees contain build output	A correctness surface of its own (nested rules, negations)
Hash-based substring where possible	Faster than Boyer-Moore on typical text	Requires little-endian + unaligned access; two code paths
mmap by default	Avoids a copy	SIGBUS on concurrent truncation
PCRE per line for regex	Full regex features with no engine work	No literal prefilter; every line pays the engine

Sources

Read at a61f1780b64266587e7bc30f0f5f71c6cca97c0f [source-verified]:

• src/util.c — boyer_moore_strnstr, hash_strnstr
• src/search.c — algorithm selection, mmap, binary skipping
• src/ignore.c, src/lang.c — ignore handling and type filters