1
0
mirror of https://github.com/golang/go synced 2024-11-19 03:04:42 -07:00
go/internal/lsp/fuzzy/matcher_test.go

296 lines
7.0 KiB
Go
Raw Normal View History

// Copyright 2019 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Benchmark results:
//
// BenchmarkMatcher-12 1000000 1615 ns/op 30.95 MB/s 0 B/op 0 allocs/op
//
package fuzzy_test
import (
"bytes"
"fmt"
"math"
"testing"
"golang.org/x/tools/internal/lsp/fuzzy"
)
type comparator struct {
f func(val, ref float32) bool
descr string
}
var (
eq = comparator{
f: func(val, ref float32) bool {
return val == ref
},
descr: "==",
}
ge = comparator{
f: func(val, ref float32) bool {
return val >= ref
},
descr: ">=",
}
internal/lsp: fix fuzzy matcher inconsistency Originally the fuzzy matcher required a match in the final candidate segment. For example, to match the candidate "foo.bar", the input had to have at least one character that matched "bar". I previously removed this requirement as it is too restrictive for deep completions to be useful. However, there was still some lingering final-segment favoritism in the matching algorithm. In particular, there were penalties for not matching the final segment's first character and for not matching the final segment's word initial characters. However, these penalties only made sense when we also required a final segment match. Consider this example: User input: "U" Candidate "ErrUnexpectedEOF" - with only a single segment, we got big penalties for not matching the leading "E" (since it is the final segment). Candidate "ErrUnexpectedEOF.Error" - "ErrUnexpectedEOF" is no longer the final segment, so we didn't get penalties. And we didn't get penalties for the final segment "Error" because we finished matching after the first "U". As a result, this candidate slips through with a higher score. Fix by simplifying the skip penalty. Now we only penalize for skipping the first character of the first or final segment (and the penalty is lower). For deep completions, the first and final segment are both "important" segments, so I think it makes sense to focus on both of them. We don't want to penalize all segment starts because that makes it harder to match deeper candidates where you often "ignore" intermediate segments. I had to adjust a few scores in the tests, but I don't think the impact will be too big other than fixing the bug. Fixes golang/go#35062. Change-Id: Id17a5c80bf0f80ce252fe990ccfbd51c1bac1c72 Reviewed-on: https://go-review.googlesource.com/c/tools/+/202638 Reviewed-by: Rebecca Stambler <rstambler@golang.org> Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2019-10-22 10:28:04 -06:00
gt = comparator{
f: func(val, ref float32) bool {
return val > ref
},
descr: ">",
}
)
func (c comparator) eval(val, ref float32) bool {
return c.f(val, ref)
}
func (c comparator) String() string {
return c.descr
}
type scoreTest struct {
candidate string
comparator
ref float32
}
var matcherTests = []struct {
pattern string
tests []scoreTest
}{
{
pattern: "",
tests: []scoreTest{
{"def", eq, 1},
{"Ab stuff c", eq, 1},
},
},
{
pattern: "abc",
tests: []scoreTest{
{"def", eq, -1},
{"abd", eq, -1},
{"abc", ge, 0},
{"Abc", ge, 0},
{"Ab stuff c", ge, 0},
},
},
{
pattern: "Abc",
tests: []scoreTest{
{"def", eq, -1},
{"abd", eq, -1},
{"abc", ge, 0},
{"Abc", ge, 0},
{"Ab stuff c", ge, 0},
},
},
internal/lsp: fix fuzzy matcher inconsistency Originally the fuzzy matcher required a match in the final candidate segment. For example, to match the candidate "foo.bar", the input had to have at least one character that matched "bar". I previously removed this requirement as it is too restrictive for deep completions to be useful. However, there was still some lingering final-segment favoritism in the matching algorithm. In particular, there were penalties for not matching the final segment's first character and for not matching the final segment's word initial characters. However, these penalties only made sense when we also required a final segment match. Consider this example: User input: "U" Candidate "ErrUnexpectedEOF" - with only a single segment, we got big penalties for not matching the leading "E" (since it is the final segment). Candidate "ErrUnexpectedEOF.Error" - "ErrUnexpectedEOF" is no longer the final segment, so we didn't get penalties. And we didn't get penalties for the final segment "Error" because we finished matching after the first "U". As a result, this candidate slips through with a higher score. Fix by simplifying the skip penalty. Now we only penalize for skipping the first character of the first or final segment (and the penalty is lower). For deep completions, the first and final segment are both "important" segments, so I think it makes sense to focus on both of them. We don't want to penalize all segment starts because that makes it harder to match deeper candidates where you often "ignore" intermediate segments. I had to adjust a few scores in the tests, but I don't think the impact will be too big other than fixing the bug. Fixes golang/go#35062. Change-Id: Id17a5c80bf0f80ce252fe990ccfbd51c1bac1c72 Reviewed-on: https://go-review.googlesource.com/c/tools/+/202638 Reviewed-by: Rebecca Stambler <rstambler@golang.org> Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2019-10-22 10:28:04 -06:00
{
pattern: "U",
tests: []scoreTest{
{"ErrUnexpectedEOF", gt, 0},
{"ErrUnexpectedEOF.Error", eq, 0},
},
},
}
func TestScore(t *testing.T) {
for _, tc := range matcherTests {
m := fuzzy.NewMatcher(tc.pattern)
for _, sct := range tc.tests {
score := m.Score(sct.candidate)
if !sct.comparator.eval(score, sct.ref) {
t.Errorf("not true that m.Score(%s)[=%v] %s %v", sct.candidate, score, sct.comparator, sct.ref)
}
}
}
}
var compareCandidatesTestCases = []struct {
pattern string
orderedCandidates []string
}{
{
pattern: "Foo",
orderedCandidates: []string{
"Barfoo",
"Faoo",
"F_o_o",
internal/lsp: fix fuzzy matcher inconsistency Originally the fuzzy matcher required a match in the final candidate segment. For example, to match the candidate "foo.bar", the input had to have at least one character that matched "bar". I previously removed this requirement as it is too restrictive for deep completions to be useful. However, there was still some lingering final-segment favoritism in the matching algorithm. In particular, there were penalties for not matching the final segment's first character and for not matching the final segment's word initial characters. However, these penalties only made sense when we also required a final segment match. Consider this example: User input: "U" Candidate "ErrUnexpectedEOF" - with only a single segment, we got big penalties for not matching the leading "E" (since it is the final segment). Candidate "ErrUnexpectedEOF.Error" - "ErrUnexpectedEOF" is no longer the final segment, so we didn't get penalties. And we didn't get penalties for the final segment "Error" because we finished matching after the first "U". As a result, this candidate slips through with a higher score. Fix by simplifying the skip penalty. Now we only penalize for skipping the first character of the first or final segment (and the penalty is lower). For deep completions, the first and final segment are both "important" segments, so I think it makes sense to focus on both of them. We don't want to penalize all segment starts because that makes it harder to match deeper candidates where you often "ignore" intermediate segments. I had to adjust a few scores in the tests, but I don't think the impact will be too big other than fixing the bug. Fixes golang/go#35062. Change-Id: Id17a5c80bf0f80ce252fe990ccfbd51c1bac1c72 Reviewed-on: https://go-review.googlesource.com/c/tools/+/202638 Reviewed-by: Rebecca Stambler <rstambler@golang.org> Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2019-10-22 10:28:04 -06:00
"FaoFooa",
"BarFoo",
"F__oo",
"F_oo",
"FooA",
"FooBar",
"Foo",
},
},
internal/lsp: fix fuzzy matcher inconsistency Originally the fuzzy matcher required a match in the final candidate segment. For example, to match the candidate "foo.bar", the input had to have at least one character that matched "bar". I previously removed this requirement as it is too restrictive for deep completions to be useful. However, there was still some lingering final-segment favoritism in the matching algorithm. In particular, there were penalties for not matching the final segment's first character and for not matching the final segment's word initial characters. However, these penalties only made sense when we also required a final segment match. Consider this example: User input: "U" Candidate "ErrUnexpectedEOF" - with only a single segment, we got big penalties for not matching the leading "E" (since it is the final segment). Candidate "ErrUnexpectedEOF.Error" - "ErrUnexpectedEOF" is no longer the final segment, so we didn't get penalties. And we didn't get penalties for the final segment "Error" because we finished matching after the first "U". As a result, this candidate slips through with a higher score. Fix by simplifying the skip penalty. Now we only penalize for skipping the first character of the first or final segment (and the penalty is lower). For deep completions, the first and final segment are both "important" segments, so I think it makes sense to focus on both of them. We don't want to penalize all segment starts because that makes it harder to match deeper candidates where you often "ignore" intermediate segments. I had to adjust a few scores in the tests, but I don't think the impact will be too big other than fixing the bug. Fixes golang/go#35062. Change-Id: Id17a5c80bf0f80ce252fe990ccfbd51c1bac1c72 Reviewed-on: https://go-review.googlesource.com/c/tools/+/202638 Reviewed-by: Rebecca Stambler <rstambler@golang.org> Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2019-10-22 10:28:04 -06:00
{
pattern: "U",
orderedCandidates: []string{
"ErrUnexpectedEOF.Error",
"ErrUnexpectedEOF",
},
},
}
func TestCompareCandidateScores(t *testing.T) {
for _, tc := range compareCandidatesTestCases {
m := fuzzy.NewMatcher(tc.pattern)
var prevScore float32
prevCand := "MIN_SCORE"
for _, cand := range tc.orderedCandidates {
score := m.Score(cand)
if prevScore > score {
t.Errorf("%s[=%v] is scored lower than %s[=%v]", cand, score, prevCand, prevScore)
}
if score < -1 || score > 1 {
t.Errorf("%s score is %v; want value between [-1, 1]", cand, score)
}
prevScore = score
prevCand = cand
}
}
}
var fuzzyMatcherTestCases = []struct {
p string
str string
want string
}{
{p: "foo", str: "abc::foo", want: "abc::[foo]"},
{p: "foo", str: "foo.foo", want: "foo.[foo]"},
{p: "foo", str: "fo_oo.o_oo", want: "[fo]_oo.[o]_oo"},
{p: "foo", str: "fo_oo.fo_oo", want: "fo_oo.[fo]_[o]o"},
{p: "fo_o", str: "fo_oo.o_oo", want: "[f]o_oo.[o_o]o"},
{p: "fOO", str: "fo_oo.o_oo", want: "[f]o_oo.[o]_[o]o"},
{p: "tedit", str: "foo.TextEdit", want: "foo.[T]ext[Edit]"},
{p: "TEdit", str: "foo.TextEdit", want: "foo.[T]ext[Edit]"},
{p: "Tedit", str: "foo.TextEdit", want: "foo.[T]ext[Edit]"},
{p: "Tedit", str: "foo.Textedit", want: "foo.[Te]xte[dit]"},
{p: "TEdit", str: "foo.Textedit", want: ""},
{p: "te", str: "foo.Textedit", want: "foo.[Te]xtedit"},
{p: "ee", str: "foo.Textedit", want: ""}, // short middle of the word match
{p: "ex", str: "foo.Textedit", want: "foo.T[ex]tedit"},
{p: "exdi", str: "foo.Textedit", want: ""}, // short middle of the word match
{p: "exdit", str: "foo.Textedit", want: ""}, // short middle of the word match
{p: "extdit", str: "foo.Textedit", want: "foo.T[ext]e[dit]"},
{p: "e", str: "foo.Textedit", want: "foo.T[e]xtedit"},
{p: "E", str: "foo.Textedit", want: "foo.T[e]xtedit"},
{p: "ed", str: "foo.Textedit", want: "foo.Text[ed]it"},
{p: "edt", str: "foo.Textedit", want: ""}, // short middle of the word match
{p: "edit", str: "foo.Textedit", want: "foo.Text[edit]"},
{p: "edin", str: "foo.TexteditNum", want: "foo.Text[edi]t[N]um"},
{p: "n", str: "node.GoNodeMax", want: "[n]ode.GoNodeMax"},
{p: "N", str: "node.GoNodeMax", want: "[n]ode.GoNodeMax"},
{p: "completio", str: "completion", want: "[completio]n"},
{p: "completio", str: "completion.None", want: "[completio]n.None"},
}
func TestFuzzyMatcherRanges(t *testing.T) {
for _, tc := range fuzzyMatcherTestCases {
matcher := fuzzy.NewMatcher(tc.p)
score := matcher.Score(tc.str)
if tc.want == "" {
if score >= 0 {
t.Errorf("Score(%s, %s) = %v; want: <= 0", tc.p, tc.str, score)
}
continue
}
if score < 0 {
t.Errorf("Score(%s, %s) = %v, want: > 0", tc.p, tc.str, score)
continue
}
got := highlightMatches(tc.str, matcher)
if tc.want != got {
t.Errorf("highlightMatches(%s, %s) = %v, want: %v", tc.p, tc.str, got, tc.want)
}
}
}
var scoreTestCases = []struct {
p string
str string
want float64
}{
// Score precision up to five digits. Modify if changing the score, but make sure the new values
// are reasonable.
{p: "abc", str: "abc", want: 1},
{p: "abc", str: "Abc", want: 1},
{p: "abc", str: "Abcdef", want: 1},
{p: "strc", str: "StrCat", want: 1},
{p: "abc_def", str: "abc_def_xyz", want: 1},
{p: "abcdef", str: "abc_def_xyz", want: 0.91667},
internal/lsp: fix fuzzy matcher inconsistency Originally the fuzzy matcher required a match in the final candidate segment. For example, to match the candidate "foo.bar", the input had to have at least one character that matched "bar". I previously removed this requirement as it is too restrictive for deep completions to be useful. However, there was still some lingering final-segment favoritism in the matching algorithm. In particular, there were penalties for not matching the final segment's first character and for not matching the final segment's word initial characters. However, these penalties only made sense when we also required a final segment match. Consider this example: User input: "U" Candidate "ErrUnexpectedEOF" - with only a single segment, we got big penalties for not matching the leading "E" (since it is the final segment). Candidate "ErrUnexpectedEOF.Error" - "ErrUnexpectedEOF" is no longer the final segment, so we didn't get penalties. And we didn't get penalties for the final segment "Error" because we finished matching after the first "U". As a result, this candidate slips through with a higher score. Fix by simplifying the skip penalty. Now we only penalize for skipping the first character of the first or final segment (and the penalty is lower). For deep completions, the first and final segment are both "important" segments, so I think it makes sense to focus on both of them. We don't want to penalize all segment starts because that makes it harder to match deeper candidates where you often "ignore" intermediate segments. I had to adjust a few scores in the tests, but I don't think the impact will be too big other than fixing the bug. Fixes golang/go#35062. Change-Id: Id17a5c80bf0f80ce252fe990ccfbd51c1bac1c72 Reviewed-on: https://go-review.googlesource.com/c/tools/+/202638 Reviewed-by: Rebecca Stambler <rstambler@golang.org> Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2019-10-22 10:28:04 -06:00
{p: "abcxyz", str: "abc_def_xyz", want: 0.91667},
{p: "sc", str: "StrCat", want: 0.75},
internal/lsp: fix fuzzy matcher inconsistency Originally the fuzzy matcher required a match in the final candidate segment. For example, to match the candidate "foo.bar", the input had to have at least one character that matched "bar". I previously removed this requirement as it is too restrictive for deep completions to be useful. However, there was still some lingering final-segment favoritism in the matching algorithm. In particular, there were penalties for not matching the final segment's first character and for not matching the final segment's word initial characters. However, these penalties only made sense when we also required a final segment match. Consider this example: User input: "U" Candidate "ErrUnexpectedEOF" - with only a single segment, we got big penalties for not matching the leading "E" (since it is the final segment). Candidate "ErrUnexpectedEOF.Error" - "ErrUnexpectedEOF" is no longer the final segment, so we didn't get penalties. And we didn't get penalties for the final segment "Error" because we finished matching after the first "U". As a result, this candidate slips through with a higher score. Fix by simplifying the skip penalty. Now we only penalize for skipping the first character of the first or final segment (and the penalty is lower). For deep completions, the first and final segment are both "important" segments, so I think it makes sense to focus on both of them. We don't want to penalize all segment starts because that makes it harder to match deeper candidates where you often "ignore" intermediate segments. I had to adjust a few scores in the tests, but I don't think the impact will be too big other than fixing the bug. Fixes golang/go#35062. Change-Id: Id17a5c80bf0f80ce252fe990ccfbd51c1bac1c72 Reviewed-on: https://go-review.googlesource.com/c/tools/+/202638 Reviewed-by: Rebecca Stambler <rstambler@golang.org> Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2019-10-22 10:28:04 -06:00
{p: "abc", str: "AbstrBasicCtor", want: 0.83333},
{p: "foo", str: "abc::foo", want: 0.91667},
{p: "afoo", str: "abc::foo", want: 0.9375},
{p: "abr", str: "abc::bar", want: 0.5},
internal/lsp: fix fuzzy matcher inconsistency Originally the fuzzy matcher required a match in the final candidate segment. For example, to match the candidate "foo.bar", the input had to have at least one character that matched "bar". I previously removed this requirement as it is too restrictive for deep completions to be useful. However, there was still some lingering final-segment favoritism in the matching algorithm. In particular, there were penalties for not matching the final segment's first character and for not matching the final segment's word initial characters. However, these penalties only made sense when we also required a final segment match. Consider this example: User input: "U" Candidate "ErrUnexpectedEOF" - with only a single segment, we got big penalties for not matching the leading "E" (since it is the final segment). Candidate "ErrUnexpectedEOF.Error" - "ErrUnexpectedEOF" is no longer the final segment, so we didn't get penalties. And we didn't get penalties for the final segment "Error" because we finished matching after the first "U". As a result, this candidate slips through with a higher score. Fix by simplifying the skip penalty. Now we only penalize for skipping the first character of the first or final segment (and the penalty is lower). For deep completions, the first and final segment are both "important" segments, so I think it makes sense to focus on both of them. We don't want to penalize all segment starts because that makes it harder to match deeper candidates where you often "ignore" intermediate segments. I had to adjust a few scores in the tests, but I don't think the impact will be too big other than fixing the bug. Fixes golang/go#35062. Change-Id: Id17a5c80bf0f80ce252fe990ccfbd51c1bac1c72 Reviewed-on: https://go-review.googlesource.com/c/tools/+/202638 Reviewed-by: Rebecca Stambler <rstambler@golang.org> Run-TryBot: Rebecca Stambler <rstambler@golang.org> TryBot-Result: Gobot Gobot <gobot@golang.org>
2019-10-22 10:28:04 -06:00
{p: "br", str: "abc::bar", want: 0.25},
{p: "aar", str: "abc::bar", want: 0.41667},
{p: "edin", str: "foo.TexteditNum", want: 0.125},
{p: "ediu", str: "foo.TexteditNum", want: 0},
// We want the next two items to have roughly similar scores.
{p: "up", str: "unique_ptr", want: 0.75},
{p: "up", str: "upper_bound", want: 1},
}
func TestScores(t *testing.T) {
for _, tc := range scoreTestCases {
matcher := fuzzy.NewMatcher(tc.p)
got := math.Round(float64(matcher.Score(tc.str))*1e5) / 1e5
if got != tc.want {
t.Errorf("Score(%s, %s) = %v, want: %v", tc.p, tc.str, got, tc.want)
}
}
}
func highlightMatches(str string, matcher *fuzzy.Matcher) string {
matches := matcher.MatchedRanges()
var buf bytes.Buffer
index := 0
for i := 0; i < len(matches)-1; i += 2 {
s, e := matches[i], matches[i+1]
fmt.Fprintf(&buf, "%s[%s]", str[index:s], str[s:e])
index = e
}
buf.WriteString(str[index:])
return buf.String()
}
func BenchmarkMatcher(b *testing.B) {
pattern := "Foo"
candidates := []string{
"F_o_o",
"Barfoo",
"Faoo",
"F__oo",
"F_oo",
"FaoFooa",
"BarFoo",
"FooA",
"FooBar",
"Foo",
}
matcher := fuzzy.NewMatcher(pattern)
b.ResetTimer()
for i := 0; i < b.N; i++ {
for _, c := range candidates {
matcher.Score(c)
}
}
var numBytes int
for _, c := range candidates {
numBytes += len(c)
}
b.SetBytes(int64(numBytes))
}