plumbing: blame, Complete rewrite. Fixes #603

Signed-off-by: Arieh Schneier <15041913+AriehSchneier@users.noreply.github.com>

1 files changed, 450 insertions, 162 deletions

diff --git a/blame.go b/blame.go
index 43634b3..2a877dc 100644
--- a/blame.go
+++ b/blame.go
@@ -2,16 +2,18 @@ package git
 
 import (
 	"bytes"
+	"container/heap"
 	"errors"
 	"fmt"
+	"io"
 	"strconv"
-	"strings"
 	"time"
 	"unicode/utf8"
 
 	"github.com/go-git/go-git/v5/plumbing"
 	"github.com/go-git/go-git/v5/plumbing/object"
 	"github.com/go-git/go-git/v5/utils/diff"
+	"github.com/sergi/go-diff/diffmatchpatch"
 )
 
 // BlameResult represents the result of a Blame operation.
@@ -29,67 +31,86 @@ type BlameResult struct {
 func Blame(c *object.Commit, path string) (*BlameResult, error) {
 	// The file to blame is identified by the input arguments:
 	// commit and path. commit is a Commit object obtained from a Repository. Path
-	// represents a path to a specific file contained into the repository.
+	// represents a path to a specific file contained in the repository.
 	//
-	// Blaming a file is a two step process:
+	// Blaming a file is done by walking the tree in reverse order trying to find where each line was last modified.
 	//
-	// 1. Create a linear history of the commits affecting a file. We use
-	// revlist.New for that.
+	// When a diff is found it cannot immediately assume it came from that commit, as it may have come from 1 of its
+	// parents, so it will first try to resolve those diffs from its parents, if it couldn't find the change in its
+	// parents then it will assign the change to itself.
 	//
-	// 2. Then build a graph with a node for every line in every file in
-	// the history of the file.
+	// When encountering 2 parents that have made the same change to a file it will choose the parent that was merged
+	// into the current branch first (this is determined by the order of the parents inside the commit).
 	//
-	// Each node is assigned a commit: Start by the nodes in the first
-	// commit. Assign that commit as the creator of all its lines.
-	//
-	// Then jump to the nodes in the next commit, and calculate the diff
-	// between the two files. Newly created lines get
-	// assigned the new commit as its origin. Modified lines also get
-	// this new commit. Untouched lines retain the old commit.
-	//
-	// All this work is done in the assignOrigin function which holds all
-	// the internal relevant data in a "blame" struct, that is not
-	// exported.
-	//
-	// TODO: ways to improve the efficiency of this function:
-	// 1. Improve revlist
-	// 2. Improve how to traverse the history (example a backward traversal will
-	// be much more efficient)
-	//
-	// TODO: ways to improve the function in general:
-	// 1. Add memoization between revlist and assign.
-	// 2. It is using much more memory than needed, see the TODOs below.
+	// This currently works on a line by line basis, if performance becomes an issue it could be changed to work with
+	// hunks rather than lines. Then when encountering diff hunks it would need to split them where necessary.
 
 	b := new(blame)
 	b.fRev = c
 	b.path = path
+	b.q = new(priorityQueue)
 
-	// get all the file revisions
-	if err := b.fillRevs(); err != nil {
+	file, err := b.fRev.File(path)
+	if err != nil {
 		return nil, err
 	}
-
-	// calculate the line tracking graph and fill in
-	// file contents in data.
-	if err := b.fillGraphAndData(); err != nil {
+	finalLines, err := file.Lines()
+	if err != nil {
 		return nil, err
 	}
+	finalLength := len(finalLines)
 
-	file, err := b.fRev.File(b.path)
+	needsMap := make([]lineMap, finalLength)
+	for i := range needsMap {
+		needsMap[i] = lineMap{i, i, nil, -1}
+	}
+	contents, err := file.Contents()
 	if err != nil {
 		return nil, err
 	}
-	finalLines, err := file.Lines()
+	b.q.Push(&queueItem{
+		nil,
+		nil,
+		c,
+		path,
+		contents,
+		needsMap,
+		0,
+		false,
+		0,
+	})
+	items := make([]*queueItem, 0)
+	for {
+		items = items[:0]
+		for {
+			if b.q.Len() == 0 {
+				return nil, errors.New("invalid state: no items left on the blame queue")
+			}
+			item := b.q.Pop()
+			items = append(items, item)
+			next := b.q.Peek()
+			if next == nil || next.Hash != item.Commit.Hash {
+				break
+			}
+		}
+		finished, err := b.addBlames(items)
+		if err != nil {
+			return nil, err
+		}
+		if finished == true {
+			break
+		}
+	}
 	if err != nil {
 		return nil, err
 	}
 
-	// Each node (line) holds the commit where it was introduced or
-	// last modified. To achieve that we use the FORWARD algorithm
-	// described in Zimmermann, et al. "Mining Version Archives for
-	// Co-changed Lines", in proceedings of the Mining Software
-	// Repositories workshop, Shanghai, May 22-23, 2006.
-	lines, err := newLines(finalLines, b.sliceGraph(len(b.graph)-1))
+	b.lineToCommit = make([]*object.Commit, finalLength)
+	for i := range needsMap {
+		b.lineToCommit[i] = needsMap[i].Commit
+	}
+
+	lines, err := newLines(finalLines, b.lineToCommit)
 	if err != nil {
 		return nil, err
 	}
@@ -105,6 +126,8 @@ func Blame(c *object.Commit, path string) (*BlameResult, error) {
 type Line struct {
 	// Author is the email address of the last author that modified the line.
 	Author string
+	// AuthorName is the name of the last author that modified the line.
+	AuthorName string
 	// Text is the original text of the line.
 	Text string
 	// Date is when the original text of the line was introduced
@@ -113,31 +136,21 @@ type Line struct {
 	Hash plumbing.Hash
 }
 
-func newLine(author, text string, date time.Time, hash plumbing.Hash) *Line {
+func newLine(author, authorName, text string, date time.Time, hash plumbing.Hash) *Line {
 	return &Line{
-		Author: author,
-		Text:   text,
-		Hash:   hash,
-		Date:   date,
+		Author:     author,
+		AuthorName: authorName,
+		Text:       text,
+		Hash:       hash,
+		Date:       date,
 	}
 }
 
 func newLines(contents []string, commits []*object.Commit) ([]*Line, error) {
-	lcontents := len(contents)
-	lcommits := len(commits)
-
-	if lcontents != lcommits {
-		if lcontents == lcommits-1 && contents[lcontents-1] != "\n" {
-			contents = append(contents, "\n")
-		} else {
-			return nil, errors.New("contents and commits have different length")
-		}
-	}
-
-	result := make([]*Line, 0, lcontents)
+	result := make([]*Line, 0, len(contents))
 	for i := range contents {
 		result = append(result, newLine(
-			commits[i].Author.Email, contents[i],
+			commits[i].Author.Email, commits[i].Author.Name, contents[i],
 			commits[i].Author.When, commits[i].Hash,
 		))
 	}
@@ -152,151 +165,426 @@ type blame struct {
 	path string
 	// the commit of the final revision of the file to blame
 	fRev *object.Commit
-	// the chain of revisions affecting the the file to blame
-	revs []*object.Commit
-	// the contents of the file across all its revisions
-	data []string
-	// the graph of the lines in the file across all the revisions
-	graph [][]*object.Commit
+	// resolved lines
+	lineToCommit []*object.Commit
+	// queue of commits that need resolving
+	q *priorityQueue
 }
 
-// calculate the history of a file "path", starting from commit "from", sorted by commit date.
-func (b *blame) fillRevs() error {
-	var err error
-
-	b.revs, err = references(b.fRev, b.path)
-	return err
+type lineMap struct {
+	Orig, Cur    int
+	Commit       *object.Commit
+	FromParentNo int
 }
 
-// build graph of a file from its revision history
-func (b *blame) fillGraphAndData() error {
-	//TODO: not all commits are needed, only the current rev and the prev
-	b.graph = make([][]*object.Commit, len(b.revs))
-	b.data = make([]string, len(b.revs)) // file contents in all the revisions
-	// for every revision of the file, starting with the first
-	// one...
-	for i, rev := range b.revs {
+func (b *blame) addBlames(curItems []*queueItem) (bool, error) {
+	curItem := curItems[0]
+
+	// Simple optimisation to merge paths, there is potential to go a bit further here and check for any duplicates
+	// not only if they are all the same.
+	if len(curItems) == 1 {
+		curItems = nil
+	} else if curItem.IdenticalToChild {
+		allSame := true
+		lenCurItems := len(curItems)
+		lowestParentNo := curItem.ParentNo
+		for i := 1; i < lenCurItems; i++ {
+			if !curItems[i].IdenticalToChild || curItem.Child != curItems[i].Child {
+				allSame = false
+				break
+			}
+			lowestParentNo = min(lowestParentNo, curItems[i].ParentNo)
+		}
+		if allSame {
+			curItem.Child.numParentsNeedResolving = curItem.Child.numParentsNeedResolving - lenCurItems + 1
+			curItems = nil // free the memory
+			curItem.ParentNo = lowestParentNo
+
+			// Now check if we can remove the parent completely
+			for curItem.Child.IdenticalToChild && curItem.Child.MergedChildren == nil && curItem.Child.numParentsNeedResolving == 1 {
+				oldChild := curItem.Child
+				curItem.Child = oldChild.Child
+				curItem.ParentNo = oldChild.ParentNo
+			}
+		}
+	}
+
+	// if we have more than 1 item for this commit, create a single needsMap
+	if len(curItems) > 1 {
+		curItem.MergedChildren = make([]childToNeedsMap, len(curItems))
+		for i, c := range curItems {
+			curItem.MergedChildren[i] = childToNeedsMap{c.Child, c.NeedsMap, c.IdenticalToChild, c.ParentNo}
+		}
+		newNeedsMap := make([]lineMap, 0, len(curItem.NeedsMap))
+		newNeedsMap = append(newNeedsMap, curItems[0].NeedsMap...)
+
+		for i := 1; i < len(curItems); i++ {
+			cur := curItems[i].NeedsMap
+			n := 0 // position in newNeedsMap
+			c := 0 // position in current list
+			for c < len(cur) {
+				if n == len(newNeedsMap) {
+					newNeedsMap = append(newNeedsMap, cur[c:]...)
+					break
+				} else if newNeedsMap[n].Cur == cur[c].Cur {
+					n++
+					c++
+				} else if newNeedsMap[n].Cur < cur[c].Cur {
+					n++
+				} else {
+					newNeedsMap = append(newNeedsMap, cur[c])
+					newPos := len(newNeedsMap) - 1
+					for newPos > n {
+						newNeedsMap[newPos-1], newNeedsMap[newPos] = newNeedsMap[newPos], newNeedsMap[newPos-1]
+						newPos--
+					}
+				}
+			}
+		}
+		curItem.NeedsMap = newNeedsMap
+		curItem.IdenticalToChild = false
+		curItem.Child = nil
+		curItems = nil // free the memory
+	}
+
+	parents, err := parentsContainingPath(curItem.path, curItem.Commit)
+	if err != nil {
+		return false, err
+	}
+
+	anyPushed := false
+	for parnetNo, prev := range parents {
+		currentHash, err := blobHash(curItem.path, curItem.Commit)
+		if err != nil {
+			return false, err
+		}
+		prevHash, err := blobHash(prev.Path, prev.Commit)
+		if err != nil {
+			return false, err
+		}
+		if currentHash == prevHash {
+			if len(parents) == 1 && curItem.MergedChildren == nil && curItem.IdenticalToChild {
+				// commit that has 1 parent and 1 child and is the same as both, bypass it completely
+				b.q.Push(&queueItem{
+					Child:            curItem.Child,
+					Commit:           prev.Commit,
+					path:             prev.Path,
+					Contents:         curItem.Contents,
+					NeedsMap:         curItem.NeedsMap, // reuse the NeedsMap as we are throwing away this item
+					IdenticalToChild: true,
+					ParentNo:         curItem.ParentNo,
+				})
+			} else {
+				b.q.Push(&queueItem{
+					Child:            curItem,
+					Commit:           prev.Commit,
+					path:             prev.Path,
+					Contents:         curItem.Contents,
+					NeedsMap:         append([]lineMap(nil), curItem.NeedsMap...), // create new slice and copy
+					IdenticalToChild: true,
+					ParentNo:         parnetNo,
+				})
+				curItem.numParentsNeedResolving++
+			}
+			anyPushed = true
+			continue
+		}
+
 		// get the contents of the file
-		file, err := rev.File(b.path)
+		file, err := prev.Commit.File(prev.Path)
 		if err != nil {
-			return nil
+			return false, err
 		}
-		b.data[i], err = file.Contents()
+		prevContents, err := file.Contents()
 		if err != nil {
-			return err
+			return false, err
 		}
-		nLines := countLines(b.data[i])
-		// create a node for each line
-		b.graph[i] = make([]*object.Commit, nLines)
-		// assign a commit to each node
-		// if this is the first revision, then the node is assigned to
-		// this first commit.
-		if i == 0 {
-			for j := 0; j < nLines; j++ {
-				b.graph[i][j] = b.revs[i]
+
+		hunks := diff.Do(prevContents, curItem.Contents)
+		prevl := -1
+		curl := -1
+		need := 0
+		getFromParent := make([]lineMap, 0)
+	out:
+		for h := range hunks {
+			hLines := countLines(hunks[h].Text)
+			for hl := 0; hl < hLines; hl++ {
+				switch {
+				case hunks[h].Type == diffmatchpatch.DiffEqual:
+					prevl++
+					curl++
+					if curl == curItem.NeedsMap[need].Cur {
+						// add to needs
+						getFromParent = append(getFromParent, lineMap{curl, prevl, nil, -1})
+						// move to next need
+						need++
+						if need >= len(curItem.NeedsMap) {
+							break out
+						}
+					}
+				case hunks[h].Type == diffmatchpatch.DiffInsert:
+					curl++
+					if curl == curItem.NeedsMap[need].Cur {
+						// the line we want is added, it may have been added here (or by another parent), skip it for now
+						need++
+						if need >= len(curItem.NeedsMap) {
+							break out
+						}
+					}
+				case hunks[h].Type == diffmatchpatch.DiffDelete:
+					prevl += hLines
+					continue out
+				default:
+					return false, errors.New("invalid state: invalid hunk Type")
+				}
 			}
-		} else {
-			// if this is not the first commit, then assign to the old
-			// commit or to the new one, depending on what the diff
-			// says.
-			b.assignOrigin(i, i-1)
+		}
+
+		if len(getFromParent) > 0 {
+			b.q.Push(&queueItem{
+				curItem,
+				nil,
+				prev.Commit,
+				prev.Path,
+				prevContents,
+				getFromParent,
+				0,
+				false,
+				parnetNo,
+			})
+			curItem.numParentsNeedResolving++
+			anyPushed = true
 		}
 	}
-	return nil
-}
 
-// sliceGraph returns a slice of commits (one per line) for a particular
-// revision of a file (0=first revision).
-func (b *blame) sliceGraph(i int) []*object.Commit {
-	fVs := b.graph[i]
-	result := make([]*object.Commit, 0, len(fVs))
-	for _, v := range fVs {
-		c := *v
-		result = append(result, &c)
+	curItem.Contents = "" // no longer need, free the memory
+
+	if !anyPushed {
+		return finishNeeds(curItem)
 	}
-	return result
+
+	return false, nil
 }
 
-// Assigns origin to vertexes in current (c) rev from data in its previous (p)
-// revision
-func (b *blame) assignOrigin(c, p int) {
-	// assign origin based on diff info
-	hunks := diff.Do(b.data[p], b.data[c])
-	sl := -1 // source line
-	dl := -1 // destination line
-	for h := range hunks {
-		hLines := countLines(hunks[h].Text)
-		for hl := 0; hl < hLines; hl++ {
-			switch {
-			case hunks[h].Type == 0:
-				sl++
-				dl++
-				b.graph[c][dl] = b.graph[p][sl]
-			case hunks[h].Type == 1:
-				dl++
-				b.graph[c][dl] = b.revs[c]
-			case hunks[h].Type == -1:
-				sl++
-			default:
-				panic("unreachable")
+func finishNeeds(curItem *queueItem) (bool, error) {
+	// any needs left in the needsMap must have come from this revision
+	for i := range curItem.NeedsMap {
+		if curItem.NeedsMap[i].Commit == nil {
+			curItem.NeedsMap[i].Commit = curItem.Commit
+			curItem.NeedsMap[i].FromParentNo = -1
+		}
+	}
+
+	if curItem.Child == nil && curItem.MergedChildren == nil {
+		return true, nil
+	}
+
+	if curItem.MergedChildren == nil {
+		return applyNeeds(curItem.Child, curItem.NeedsMap, curItem.IdenticalToChild, curItem.ParentNo)
+	}
+
+	for _, ctn := range curItem.MergedChildren {
+		m := 0 // position in merged needs map
+		p := 0 // position in parent needs map
+		for p < len(ctn.NeedsMap) {
+			if ctn.NeedsMap[p].Cur == curItem.NeedsMap[m].Cur {
+				ctn.NeedsMap[p].Commit = curItem.NeedsMap[m].Commit
+				m++
+				p++
+			} else if ctn.NeedsMap[p].Cur < curItem.NeedsMap[m].Cur {
+				p++
+			} else {
+				m++
 			}
 		}
+		finished, err := applyNeeds(ctn.Child, ctn.NeedsMap, ctn.IdenticalToChild, ctn.ParentNo)
+		if finished || err != nil {
+			return finished, err
+		}
 	}
-}
 
-// GoString prints the results of a Blame using git-blame's style.
-func (b *blame) GoString() string {
-	var buf bytes.Buffer
+	return false, nil
+}
 
-	file, err := b.fRev.File(b.path)
-	if err != nil {
-		panic("PrettyPrint: internal error in repo.Data")
+func applyNeeds(child *queueItem, needsMap []lineMap, identicalToChild bool, parentNo int) (bool, error) {
+	if identicalToChild {
+		for i := range child.NeedsMap {
+			l := &child.NeedsMap[i]
+			if l.Cur != needsMap[i].Cur || l.Orig != needsMap[i].Orig {
+				return false, errors.New("needsMap isn't the same? Why not??")
+			}
+			if l.Commit == nil || parentNo < l.FromParentNo {
+				l.Commit = needsMap[i].Commit
+				l.FromParentNo = parentNo
+			}
+		}
+	} else {
+		i := 0
+	out:
+		for j := range child.NeedsMap {
+			l := &child.NeedsMap[j]
+			for needsMap[i].Orig < l.Cur {
+				i++
+				if i == len(needsMap) {
+					break out
+				}
+			}
+			if l.Cur == needsMap[i].Orig {
+				if l.Commit == nil || parentNo < l.FromParentNo {
+					l.Commit = needsMap[i].Commit
+					l.FromParentNo = parentNo
+				}
+			}
+		}
 	}
-	contents, err := file.Contents()
-	if err != nil {
-		panic("PrettyPrint: internal error in repo.Data")
+	child.numParentsNeedResolving--
+	if child.numParentsNeedResolving == 0 {
+		finished, err := finishNeeds(child)
+		if finished || err != nil {
+			return finished, err
+		}
 	}
 
-	lines := strings.Split(contents, "\n")
+	return false, nil
+}
+
+// String prints the results of a Blame using git-blame's style.
+func (b BlameResult) String() string {
+	var buf bytes.Buffer
+
 	// max line number length
-	mlnl := len(strconv.Itoa(len(lines)))
+	mlnl := len(strconv.Itoa(len(b.Lines)))
 	// max author length
 	mal := b.maxAuthorLength()
-	format := fmt.Sprintf("%%s (%%-%ds %%%dd) %%s\n",
-		mal, mlnl)
+	format := fmt.Sprintf("%%s (%%-%ds %%s %%%dd) %%s\n", mal, mlnl)
 
-	fVs := b.graph[len(b.graph)-1]
-	for ln, v := range fVs {
-		fmt.Fprintf(&buf, format, v.Hash.String()[:8],
-			prettyPrintAuthor(fVs[ln]), ln+1, lines[ln])
+	for ln := range b.Lines {
+		_, _ = fmt.Fprintf(&buf, format, b.Lines[ln].Hash.String()[:8],
+			b.Lines[ln].AuthorName, b.Lines[ln].Date.Format("2006-01-02 15:04:05 -0700"), ln+1, b.Lines[ln].Text)
 	}
 	return buf.String()
 }
 
-// utility function to pretty print the author.
-func prettyPrintAuthor(c *object.Commit) string {
-	return fmt.Sprintf("%s %s", c.Author.Name, c.Author.When.Format("2006-01-02"))
-}
-
 // utility function to calculate the number of runes needed
 // to print the longest author name in the blame of a file.
-func (b *blame) maxAuthorLength() int {
-	memo := make(map[plumbing.Hash]struct{}, len(b.graph)-1)
-	fVs := b.graph[len(b.graph)-1]
+func (b BlameResult) maxAuthorLength() int {
 	m := 0
-	for ln := range fVs {
-		if _, ok := memo[fVs[ln].Hash]; ok {
-			continue
-		}
-		memo[fVs[ln].Hash] = struct{}{}
-		m = max(m, utf8.RuneCountInString(prettyPrintAuthor(fVs[ln])))
+	for ln := range b.Lines {
+		m = max(m, utf8.RuneCountInString(b.Lines[ln].AuthorName))
 	}
 	return m
 }
 
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
 func max(a, b int) int {
 	if a > b {
 		return a
 	}
 	return b
 }
+
+type childToNeedsMap struct {
+	Child            *queueItem
+	NeedsMap         []lineMap
+	IdenticalToChild bool
+	ParentNo         int
+}
+
+type queueItem struct {
+	Child                   *queueItem
+	MergedChildren          []childToNeedsMap
+	Commit                  *object.Commit
+	path                    string
+	Contents                string
+	NeedsMap                []lineMap
+	numParentsNeedResolving int
+	IdenticalToChild        bool
+	ParentNo                int
+}
+
+type priorityQueueImp []*queueItem
+
+func (pq *priorityQueueImp) Len() int { return len(*pq) }
+func (pq *priorityQueueImp) Less(i, j int) bool {
+	return !(*pq)[i].Commit.Less((*pq)[j].Commit)
+}
+func (pq *priorityQueueImp) Swap(i, j int) { (*pq)[i], (*pq)[j] = (*pq)[j], (*pq)[i] }
+func (pq *priorityQueueImp) Push(x any)    { *pq = append(*pq, x.(*queueItem)) }
+func (pq *priorityQueueImp) Pop() any {
+	n := len(*pq)
+	ret := (*pq)[n-1]
+	(*pq)[n-1] = nil // ovoid memory leak
+	*pq = (*pq)[0 : n-1]
+
+	return ret
+}
+func (pq *priorityQueueImp) Peek() *object.Commit {
+	if len(*pq) == 0 {
+		return nil
+	}
+	return (*pq)[0].Commit
+}
+
+type priorityQueue priorityQueueImp
+
+func (pq *priorityQueue) Init()    { heap.Init((*priorityQueueImp)(pq)) }
+func (pq *priorityQueue) Len() int { return (*priorityQueueImp)(pq).Len() }
+func (pq *priorityQueue) Push(c *queueItem) {
+	heap.Push((*priorityQueueImp)(pq), c)
+}
+func (pq *priorityQueue) Pop() *queueItem {
+	return heap.Pop((*priorityQueueImp)(pq)).(*queueItem)
+}
+func (pq *priorityQueue) Peek() *object.Commit { return (*priorityQueueImp)(pq).Peek() }
+
+type parentCommit struct {
+	Commit *object.Commit
+	Path   string
+}
+
+func parentsContainingPath(path string, c *object.Commit) ([]parentCommit, error) {
+	// TODO: benchmark this method making git.object.Commit.parent public instead of using
+	// an iterator
+	var result []parentCommit
+	iter := c.Parents()
+	for {
+		parent, err := iter.Next()
+		if err == io.EOF {
+			return result, nil
+		}
+		if err != nil {
+			return nil, err
+		}
+		if _, err := parent.File(path); err == nil {
+			result = append(result, parentCommit{parent, path})
+		} else {
+			// look for renames
+			patch, err := parent.Patch(c)
+			if err != nil {
+				return nil, err
+			} else if patch != nil {
+				for _, fp := range patch.FilePatches() {
+					from, to := fp.Files()
+					if from != nil && to != nil && to.Path() == path {
+						result = append(result, parentCommit{parent, from.Path()})
+						break
+					}
+				}
+			}
+		}
+	}
+}
+
+func blobHash(path string, commit *object.Commit) (plumbing.Hash, error) {
+	file, err := commit.File(path)
+	if err != nil {
+		return plumbing.ZeroHash, err
+	}
+	return file.Hash, nil
+}