Diffs and blame

Git stores snapshots, not changes. Every diff you see is computed on demand by comparing two snapshots, first as trees and then line by line.

Tree diff

A tree diff compares two trees entry by entry. Tree entries are sorted by name, so Git walks both lists together like a merge of two sorted lists. Where both trees have an entry with the same ID, the contents are identical, and for a subdirectory Git skips it entirely without reading it. That shortcut is why git diff HEAD~1 HEAD is fast even in a repository with millions of files: only the directories on the path to a changed file are read.

The result is a list of added, deleted and modified paths, which git diff --raw prints.

girtSupportedAPI docs

Rename and copy detection

A raw tree diff sees a moved file as “deleted a.rs, added b.rs”. Rename detection pairs deleted and added files by how similar their contents are and reports a rename with a similarity score, R087 a.rs b.rs for 87% similar. The default threshold is 50% (-M50%). Identical contents are found first by comparing IDs, which is cheap.

Copy detection (-C) also compares added files with files that still exist, to report “b.rs was copied from a.rs”. It costs more because there are more candidates.

The cost grows with the number of candidate pairs, so Git caps it with diff.renameLimit and falls back to plain adds and deletes beyond it.

girtPartialThe caller sets the rename limit; diff.renameLimit isn’t read.issue #562

Content diff

A content diff compares two versions of a file line by line and finds a short set of inserted and deleted lines. Git offers several algorithms:

• Myers, the default, finds a minimal edit.
• Patience first matches lines that appear exactly once in each version, which keeps unrelated braces and blank lines from being paired.
• Histogram extends patience and is often the most readable (diff.algorithm=histogram).

The output is the unified diff format: hunks of changes with a few unchanged context lines around them, each headed by @@ -10,4 +10,6 @@, meaning “four lines starting at line 10 in the old version became six lines starting at line 10 in the new one”.

girtPartialMyers with unified output; no patience or histogram.issue #502

Diff attributes

Attributes in .gitattributes can change how a path is diffed:

• -diff or binary marks a path as binary, so Git prints “Binary files differ” instead of lines.
• diff=rust chooses a diff driver. A driver’s xfuncname regex finds the enclosing function for each hunk header, @@ -10,4 +10,6 @@ fn parse(input: &str). Git has built-in drivers for many languages; others are configured as diff.<driver>.xfuncname.
• A driver’s textconv command converts the file to text before diffing, for example to diff the text of PDFs or the metadata of images.

girtPartialPatches apply binary flags and xfuncname; external commands are refused. Built-in driver patterns are incomplete (#460), and some POSIX patterns differ (#461); blame ignores diff attributes.issue #462

Blame

git blame finds, for each line of a file, the commit that last changed it. It walks history backwards from the starting commit. At each commit it diffs the file against the parent’s version: lines unchanged from the parent are passed on to the parent to keep looking, and lines the commit changed are assigned to it. The walk ends when every line has an owner.

With -M, blame follows lines moved within the file; with -C, lines moved or copied from other files. --ignore-rev skips commits such as mass reformatting.

girtPartialFollows renames; no move or copy detection, ignore-revs, textconv or incremental output.issue #503