fork of https://github.com/sourcegraph/zoekt
scoring: reduce allocations for addScore (#670)
We add a bit more complexity to addScore but avoid allocating a string if debugScore = false.
Test plan:
I ran a couple of benchmarks and confirmed we don't allocate.
author
Stefan Hengl
committer
GitHub
date
(Oct 23, 2023, 1:39 PM +0200)
commit
0a17ccb2
0a17ccb25ade68111340b5faa594a0a7d98e97e4
parent
5bbf05d4
5bbf05d4365ed16c7b673d78a5b818e930288a32
+23
-9
+23
-9
eval.go
+23
-9
eval.go
···
21
21
"math"
22
22
"regexp/syntax"
23
23
"sort"
24
+
"strconv"
24
25
"strings"
25
26
"time"
26
27
···
32
33
33
34
const maxUInt16 = 0xffff
34
35
35
-
func (m *FileMatch) addScore(what string, s float64, debugScore bool) {
36
-
if s != 0 && debugScore {
37
-
m.Debug += fmt.Sprintf("%s:%.2f, ", what, s)
36
+
// addScore increments the score of the FileMatch by the computed score. If
37
+
// debugScore is true, it also adds a debug string to the FileMatch. If raw is
38
+
// -1, it is ignored. Otherwise, it is added to the debug string.
39
+
func (m *FileMatch) addScore(what string, computed float64, raw float64, debugScore bool) {
40
+
if computed != 0 && debugScore {
41
+
var b strings.Builder
42
+
fmt.Fprintf(&b, "%s", what)
43
+
if raw != -1 {
44
+
fmt.Fprintf(&b, "(%s)", strconv.FormatFloat(raw, 'f', -1, 64))
45
+
}
46
+
fmt.Fprintf(&b, ":%.2f, ", computed)
47
+
m.Debug += b.String()
38
48
}
39
-
m.Score += s
49
+
m.Score += computed
40
50
}
41
51
42
52
func (m *FileMatch) addKeywordScore(score float64, sumTf float64, L float64, debugScore bool) {
···
408
418
atomMatchCount++
409
419
})
410
420
421
+
addScore := func(what string, computed float64) {
422
+
fileMatch.addScore(what, computed, -1, opts.DebugScore)
423
+
}
424
+
411
425
// atom-count boosts files with matches from more than 1 atom. The
412
426
// maximum boost is scoreFactorAtomMatch.
413
427
if atomMatchCount > 0 {
414
-
fileMatch.addScore(fmt.Sprintf("atom(%d)", atomMatchCount), (1.0-1.0/float64(atomMatchCount))*scoreFactorAtomMatch, opts.DebugScore)
428
+
fileMatch.addScore("atom", (1.0-1.0/float64(atomMatchCount))*scoreFactorAtomMatch, float64(atomMatchCount), opts.DebugScore)
415
429
}
416
430
417
431
maxFileScore := 0.0
···
436
450
// Maintain ordering of input files. This
437
451
// strictly dominates the in-file ordering of
438
452
// the matches.
439
-
fileMatch.addScore("fragment", maxFileScore, opts.DebugScore)
453
+
addScore("fragment", maxFileScore)
440
454
441
455
if opts.UseDocumentRanks && len(d.ranks) > int(doc) {
442
456
weight := scoreFileRankFactor
···
452
466
// The file rank represents a log (base 2) count. The log ranks should be bounded at 32, but we
453
467
// cap it just in case to ensure it falls in the range [0, 1].
454
468
normalized := math.Min(1.0, ranks[0]/32.0)
455
-
fileMatch.addScore("file-rank", weight*normalized, opts.DebugScore)
469
+
addScore("file-rank", weight*normalized)
456
470
}
457
471
}
458
472
459
473
md := d.repoMetaData[d.repos[doc]]
460
-
fileMatch.addScore("doc-order", scoreFileOrderFactor*(1.0-float64(doc)/float64(len(d.boundaries))), opts.DebugScore)
461
-
fileMatch.addScore("repo-rank", scoreRepoRankFactor*float64(md.Rank)/maxUInt16, opts.DebugScore)
474
+
addScore("doc-order", scoreFileOrderFactor*(1.0-float64(doc)/float64(len(d.boundaries))))
475
+
addScore("repo-rank", scoreRepoRankFactor*float64(md.Rank)/maxUInt16)
462
476
463
477
if opts.DebugScore {
464
478
fileMatch.Debug = strings.TrimSuffix(fileMatch.Debug, ", ")