// Copyright 2014 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. // +build ignore // This program generates the trie for casing operations. The Unicode casing // algorithm requires the lookup of various properties and mappings for each // rune. The table generated by this generator combines several of the most // frequently used of these into a single trie so that they can be accessed // with a single lookup. package main import ( "bytes" "fmt" "io" "io/ioutil" "log" "reflect" "strconv" "strings" "unicode" "golang.org/x/text/internal/gen" "golang.org/x/text/internal/triegen" "golang.org/x/text/internal/ucd" "golang.org/x/text/unicode/norm" ) func main() { gen.Init() genTables() genTablesTest() gen.Repackage("gen_trieval.go", "trieval.go", "cases") } // runeInfo contains all information for a rune that we care about for casing // operations. type runeInfo struct { Rune rune entry info // trie value for this rune. CaseMode info // Simple case mappings. Simple [1 + maxCaseMode][]rune // Special casing HasSpecial bool Conditional bool Special [1 + maxCaseMode][]rune // Folding FoldSimple rune FoldSpecial rune FoldFull []rune // TODO: FC_NFKC, or equivalent data. // Properties SoftDotted bool CaseIgnorable bool Cased bool DecomposeGreek bool BreakType string BreakCat breakCategory // We care mostly about 0, Above, and IotaSubscript. CCC byte } type breakCategory int const ( breakBreak breakCategory = iota breakLetter breakIgnored ) // mapping returns the case mapping for the given case type. func (r *runeInfo) mapping(c info) string { if r.HasSpecial { return string(r.Special[c]) } if len(r.Simple[c]) != 0 { return string(r.Simple[c]) } return string(r.Rune) } func parse(file string, f func(p *ucd.Parser)) { ucd.Parse(gen.OpenUCDFile(file), f) } func parseUCD() []runeInfo { chars := make([]runeInfo, unicode.MaxRune) get := func(r rune) *runeInfo { c := &chars[r] c.Rune = r return c } parse("UnicodeData.txt", func(p *ucd.Parser) { ri := get(p.Rune(0)) ri.CCC = byte(p.Int(ucd.CanonicalCombiningClass)) ri.Simple[cLower] = p.Runes(ucd.SimpleLowercaseMapping) ri.Simple[cUpper] = p.Runes(ucd.SimpleUppercaseMapping) ri.Simple[cTitle] = p.Runes(ucd.SimpleTitlecaseMapping) if p.String(ucd.GeneralCategory) == "Lt" { ri.CaseMode = cTitle } }) // ; parse("PropList.txt", func(p *ucd.Parser) { if p.String(1) == "Soft_Dotted" { chars[p.Rune(0)].SoftDotted = true } }) // ; parse("DerivedCoreProperties.txt", func(p *ucd.Parser) { ri := get(p.Rune(0)) switch p.String(1) { case "Case_Ignorable": ri.CaseIgnorable = true case "Cased": ri.Cased = true case "Lowercase": ri.CaseMode = cLower case "Uppercase": ri.CaseMode = cUpper } }) // ; ; ; <upper> ; (<condition_list> ;)? parse("SpecialCasing.txt", func(p *ucd.Parser) { // We drop all conditional special casing and deal with them manually in // the language-specific case mappers. Rune 0x03A3 is the only one with // a conditional formatting that is not language-specific. However, // dealing with this letter is tricky, especially in a streaming // context, so we deal with it in the Caser for Greek specifically. ri := get(p.Rune(0)) if p.String(4) == "" { ri.HasSpecial = true ri.Special[cLower] = p.Runes(1) ri.Special[cTitle] = p.Runes(2) ri.Special[cUpper] = p.Runes(3) } else { ri.Conditional = true } }) // TODO: Use text breaking according to UAX #29. // <code>; <word break type> parse("auxiliary/WordBreakProperty.txt", func(p *ucd.Parser) { ri := get(p.Rune(0)) ri.BreakType = p.String(1) // We collapse the word breaking properties onto the categories we need. switch p.String(1) { // TODO: officially we need to canonicalize. case "Format", "MidLetter", "MidNumLet", "Single_Quote": ri.BreakCat = breakIgnored case "ALetter", "Hebrew_Letter", "Numeric", "Extend", "ExtendNumLet": ri.BreakCat = breakLetter } }) // <code>; <type>; <mapping> parse("CaseFolding.txt", func(p *ucd.Parser) { ri := get(p.Rune(0)) switch p.String(1) { case "C": ri.FoldSimple = p.Rune(2) ri.FoldFull = p.Runes(2) case "S": ri.FoldSimple = p.Rune(2) case "T": ri.FoldSpecial = p.Rune(2) case "F": ri.FoldFull = p.Runes(2) default: log.Fatalf("%U: unknown type: %s", p.Rune(0), p.String(1)) } }) return chars } func genTables() { chars := parseUCD() verifyProperties(chars) t := triegen.NewTrie("case") for i := range chars { c := &chars[i] makeEntry(c) t.Insert(rune(i), uint64(c.entry)) } w := gen.NewCodeWriter() defer w.WriteGoFile("tables.go", "cases") gen.WriteUnicodeVersion(w) // TODO: write CLDR version after adding a mechanism to detect that the // tables on which the manually created locale-sensitive casing code is // based hasn't changed. w.WriteVar("xorData", string(xorData)) w.WriteVar("exceptions", string(exceptionData)) sz, err := t.Gen(w, triegen.Compact(&sparseCompacter{})) if err != nil { log.Fatal(err) } w.Size += sz } func makeEntry(ri *runeInfo) { if ri.CaseIgnorable { if ri.Cased { ri.entry = cIgnorableCased } else { ri.entry = cIgnorableUncased } } else { ri.entry = ri.CaseMode } // TODO: handle soft-dotted. ccc := cccOther switch ri.CCC { case 0: // Not_Reordered ccc = cccZero case above: // Above ccc = cccAbove } if ri.BreakCat == breakBreak { ccc = cccBreak } ri.entry |= ccc if ri.CaseMode == cUncased { return } // Need to do something special. if ri.CaseMode == cTitle || ri.HasSpecial || ri.mapping(cTitle) != ri.mapping(cUpper) { makeException(ri) return } if f := string(ri.FoldFull); len(f) > 0 && f != ri.mapping(cUpper) && f != ri.mapping(cLower) { makeException(ri) return } // Rune is either lowercase or uppercase. orig := string(ri.Rune) mapped := "" if ri.CaseMode == cUpper { mapped = ri.mapping(cLower) } else { mapped = ri.mapping(cUpper) } if len(orig) != len(mapped) { makeException(ri) return } if string(ri.FoldFull) == ri.mapping(cUpper) { ri.entry |= inverseFoldBit } n := len(orig) // Create per-byte XOR mask. var b []byte for i := 0; i < n; i++ { b = append(b, orig[i]^mapped[i]) } // Remove leading 0 bytes, but keep at least one byte. for ; len(b) > 1 && b[0] == 0; b = b[1:] { } if len(b) == 1 && b[0]&0xc0 == 0 { ri.entry |= info(b[0]) << xorShift return } key := string(b) x, ok := xorCache[key] if !ok { xorData = append(xorData, 0) // for detecting start of sequence xorData = append(xorData, b...) x = len(xorData) - 1 xorCache[key] = x } ri.entry |= info(x<<xorShift) | xorIndexBit } var xorCache = map[string]int{} // xorData contains byte-wise XOR data for the least significant bytes of a // UTF-8 encoded rune. An index points to the last byte. The sequence starts // with a zero terminator. var xorData = []byte{} // See the comments in gen_trieval.go re "the exceptions slice". var exceptionData = []byte{0} // makeException encodes case mappings that cannot be expressed in a simple // XOR diff. func makeException(ri *runeInfo) { ccc := ri.entry & cccMask // Set exception bit and retain case type. ri.entry &= 0x0007 ri.entry |= exceptionBit if len(exceptionData) >= 1<<numExceptionBits { log.Fatalf("%U:exceptionData too large %x > %d bits", ri.Rune, len(exceptionData), numExceptionBits) } // Set the offset in the exceptionData array. ri.entry |= info(len(exceptionData) << exceptionShift) orig := string(ri.Rune) tc := ri.mapping(cTitle) uc := ri.mapping(cUpper) lc := ri.mapping(cLower) ff := string(ri.FoldFull) // addString sets the length of a string and adds it to the expansions array. addString := func(s string, b *byte) { if len(s) == 0 { // Zero-length mappings exist, but only for conditional casing, // which we are representing outside of this table. log.Fatalf("%U: has zero-length mapping.", ri.Rune) } *b <<= 3 if s != orig { n := len(s) if n > 7 { log.Fatalf("%U: mapping larger than 7 (%d)", ri.Rune, n) } *b |= byte(n) exceptionData = append(exceptionData, s...) } } // byte 0: exceptionData = append(exceptionData, byte(ccc)|byte(len(ff))) // byte 1: p := len(exceptionData) exceptionData = append(exceptionData, 0) if len(ff) > 7 { // May be zero-length. log.Fatalf("%U: fold string larger than 7 (%d)", ri.Rune, len(ff)) } exceptionData = append(exceptionData, ff...) ct := ri.CaseMode if ct != cLower { addString(lc, &exceptionData[p]) } if ct != cUpper { addString(uc, &exceptionData[p]) } if ct != cTitle { // If title is the same as upper, we set it to the original string so // that it will be marked as not present. This implies title case is // the same as upper case. if tc == uc { tc = orig } addString(tc, &exceptionData[p]) } } // sparseCompacter is a trie value block Compacter. There are many cases where // successive runes alternate between lower- and upper-case. This Compacter // exploits this by adding a special case type where the case value is obtained // from or-ing it with the least-significant bit of the rune, creating large // ranges of equal case values that compress well. type sparseCompacter struct { sparseBlocks [][]uint16 sparseOffsets []uint16 sparseCount int } // makeSparse returns the number of elements that compact block would contain // as well as the modified values. func makeSparse(vals []uint64) ([]uint16, int) { // Copy the values. values := make([]uint16, len(vals)) for i, v := range vals { values[i] = uint16(v) } alt := func(i int, v uint16) uint16 { if cm := info(v & fullCasedMask); cm == cUpper || cm == cLower { // Convert cLower or cUpper to cXORCase value, which has the form 11x. xor := v xor &^= 1 xor |= uint16(i&1) ^ (v & 1) xor |= 0x4 return xor } return v } var count int var previous uint16 for i, v := range values { if v != 0 { // Try if the unmodified value is equal to the previous. if v == previous { continue } // Try if the xor-ed value is equal to the previous value. a := alt(i, v) if a == previous { values[i] = a continue } // This is a new value. count++ // Use the xor-ed value if it will be identical to the next value. if p := i + 1; p < len(values) && alt(p, values[p]) == a { values[i] = a v = a } } previous = v } return values, count } func (s *sparseCompacter) Size(v []uint64) (int, bool) { _, n := makeSparse(v) // We limit using this method to having 16 entries. if n > 16 { return 0, false } return 2 + int(reflect.TypeOf(valueRange{}).Size())*n, true } func (s *sparseCompacter) Store(v []uint64) uint32 { h := uint32(len(s.sparseOffsets)) values, sz := makeSparse(v) s.sparseBlocks = append(s.sparseBlocks, values) s.sparseOffsets = append(s.sparseOffsets, uint16(s.sparseCount)) s.sparseCount += sz return h } func (s *sparseCompacter) Handler() string { // The sparse global variable and its lookup method is defined in gen_trieval.go. return "sparse.lookup" } func (s *sparseCompacter) Print(w io.Writer) (retErr error) { p := func(format string, args ...interface{}) { _, err := fmt.Fprintf(w, format, args...) if retErr == nil && err != nil { retErr = err } } ls := len(s.sparseBlocks) if ls == len(s.sparseOffsets) { s.sparseOffsets = append(s.sparseOffsets, uint16(s.sparseCount)) } p("// sparseOffsets: %d entries, %d bytes\n", ls+1, (ls+1)*2) p("var sparseOffsets = %#v\n\n", s.sparseOffsets) ns := s.sparseCount p("// sparseValues: %d entries, %d bytes\n", ns, ns*4) p("var sparseValues = [%d]valueRange {", ns) for i, values := range s.sparseBlocks { p("\n// Block %#x, offset %#x", i, s.sparseOffsets[i]) var v uint16 for i, nv := range values { if nv != v { if v != 0 { p(",hi:%#02x},", 0x80+i-1) } if nv != 0 { p("\n{value:%#04x,lo:%#02x", nv, 0x80+i) } } v = nv } if v != 0 { p(",hi:%#02x},", 0x80+len(values)-1) } } p("\n}\n\n") return } // verifyProperties that properties of the runes that are relied upon in the // implementation. Each property is marked with an identifier that is referred // to in the places where it is used. func verifyProperties(chars []runeInfo) { for i, c := range chars { r := rune(i) // Rune properties. // A.1: modifier never changes on lowercase. [ltLower] if c.CCC > 0 && unicode.ToLower(r) != r { log.Fatalf("%U: non-starter changes when lowercased", r) } // A.2: properties of decompositions starting with I or J. [ltLower] d := norm.NFD.PropertiesString(string(r)).Decomposition() if len(d) > 0 { if d[0] == 'I' || d[0] == 'J' { // A.2.1: we expect at least an ASCII character and a modifier. if len(d) < 3 { log.Fatalf("%U: length of decomposition was %d; want >= 3", r, len(d)) } // All subsequent runes are modifiers and all have the same CCC. runes := []rune(string(d[1:])) ccc := chars[runes[0]].CCC for _, mr := range runes[1:] { mc := chars[mr] // A.2.2: all modifiers have a CCC of Above or less. if ccc == 0 || ccc > above { log.Fatalf("%U: CCC of successive rune (%U) was %d; want (0,230]", r, mr, ccc) } // A.2.3: a sequence of modifiers all have the same CCC. if mc.CCC != ccc { log.Fatalf("%U: CCC of follow-up modifier (%U) was %d; want %d", r, mr, mc.CCC, ccc) } // A.2.4: for each trailing r, r in [0x300, 0x311] <=> CCC == Above. if (ccc == above) != (0x300 <= mr && mr <= 0x311) { log.Fatalf("%U: modifier %U in [U+0300, U+0311] != ccc(%U) == 230", r, mr, mr) } if i += len(string(mr)); i >= len(d) { break } } } } // A.3: no U+0307 in decomposition of Soft-Dotted rune. [ltUpper] if unicode.Is(unicode.Soft_Dotted, r) && strings.Contains(string(d), "\u0307") { log.Fatalf("%U: decomposition of soft-dotted rune may not contain U+0307", r) } // A.4: only rune U+0345 may be of CCC Iota_Subscript. [elUpper] if c.CCC == iotaSubscript && r != 0x0345 { log.Fatalf("%U: only rune U+0345 may have CCC Iota_Subscript", r) } // A.5: soft-dotted runes do not have exceptions. if c.SoftDotted && c.entry&exceptionBit != 0 { log.Fatalf("%U: soft-dotted has exception", r) } // A.6: Greek decomposition. [elUpper] if unicode.Is(unicode.Greek, r) { if b := norm.NFD.PropertiesString(string(r)).Decomposition(); b != nil { runes := []rune(string(b)) // A.6.1: If a Greek rune decomposes and the first rune of the // decomposition is greater than U+00FF, the rune is always // great and not a modifier. if f := runes[0]; unicode.IsMark(f) || f > 0xFF && !unicode.Is(unicode.Greek, f) { log.Fatalf("%U: expeced first rune of Greek decomposition to be letter, found %U", r, f) } // A.6.2: Any follow-up rune in a Greek decomposition is a // modifier of which the first should be gobbled in // decomposition. for _, m := range runes[1:] { switch m { case 0x0313, 0x0314, 0x0301, 0x0300, 0x0306, 0x0342, 0x0308, 0x0304, 0x345: default: log.Fatalf("%U: modifier %U is outside of expeced Greek modifier set", r, m) } } } } // Breaking properties. // B.1: all runes with CCC > 0 are of break type Extend. if c.CCC > 0 && c.BreakType != "Extend" { log.Fatalf("%U: CCC == %d, but got break type %s; want Extend", r, c.CCC, c.BreakType) } // B.2: all cased runes with c.CCC == 0 are of break type ALetter. if c.CCC == 0 && c.Cased && c.BreakType != "ALetter" { log.Fatalf("%U: cased, but got break type %s; want ALetter", r, c.BreakType) } // B.3: letter category. if c.CCC == 0 && c.BreakCat != breakBreak && !c.CaseIgnorable { if c.BreakCat != breakLetter { log.Fatalf("%U: check for letter break type gave %d; want %d", r, c.BreakCat, breakLetter) } } } } func genTablesTest() { w := &bytes.Buffer{} fmt.Fprintln(w, "var (") printProperties(w, "DerivedCoreProperties.txt", "Case_Ignorable", verifyIgnore) // We discard the output as we know we have perfect functions. We run them // just to verify the properties are correct. n := printProperties(ioutil.Discard, "DerivedCoreProperties.txt", "Cased", verifyCased) n += printProperties(ioutil.Discard, "DerivedCoreProperties.txt", "Lowercase", verifyLower) n += printProperties(ioutil.Discard, "DerivedCoreProperties.txt", "Uppercase", verifyUpper) if n > 0 { log.Fatalf("One of the discarded properties does not have a perfect filter.") } // <code>; <lower> ; <title> ; <upper> ; (<condition_list> ;)? fmt.Fprintln(w, "\tspecial = map[rune]struct{ toLower, toTitle, toUpper string }{") parse("SpecialCasing.txt", func(p *ucd.Parser) { // Skip conditional entries. if p.String(4) != "" { return } r := p.Rune(0) fmt.Fprintf(w, "\t\t0x%04x: {%q, %q, %q},\n", r, string(p.Runes(1)), string(p.Runes(2)), string(p.Runes(3))) }) fmt.Fprint(w, "\t}\n\n") // <code>; <type>; <runes> table := map[rune]struct{ simple, full, special string }{} parse("CaseFolding.txt", func(p *ucd.Parser) { r := p.Rune(0) t := p.String(1) v := string(p.Runes(2)) if t != "T" && v == string(unicode.ToLower(r)) { return } x := table[r] switch t { case "C": x.full = v x.simple = v case "S": x.simple = v case "F": x.full = v case "T": x.special = v } table[r] = x }) fmt.Fprintln(w, "\tfoldMap = map[rune]struct{ simple, full, special string }{") for r := rune(0); r < 0x10FFFF; r++ { x, ok := table[r] if !ok { continue } fmt.Fprintf(w, "\t\t0x%04x: {%q, %q, %q},\n", r, x.simple, x.full, x.special) } fmt.Fprint(w, "\t}\n\n") // Break property notBreak := map[rune]bool{} parse("auxiliary/WordBreakProperty.txt", func(p *ucd.Parser) { switch p.String(1) { case "Extend", "Format", "MidLetter", "MidNumLet", "Single_Quote", "ALetter", "Hebrew_Letter", "Numeric", "ExtendNumLet": notBreak[p.Rune(0)] = true } }) fmt.Fprintln(w, "\tbreakProp = []struct{ lo, hi rune }{") inBreak := false for r := rune(0); r <= lastRuneForTesting; r++ { if isBreak := !notBreak[r]; isBreak != inBreak { if isBreak { fmt.Fprintf(w, "\t\t{0x%x, ", r) } else { fmt.Fprintf(w, "0x%x},\n", r-1) } inBreak = isBreak } } if inBreak { fmt.Fprintf(w, "0x%x},\n", lastRuneForTesting) } fmt.Fprint(w, "\t}\n\n") // Word break test // Filter out all samples that do not contain cased characters. cased := map[rune]bool{} parse("DerivedCoreProperties.txt", func(p *ucd.Parser) { if p.String(1) == "Cased" { cased[p.Rune(0)] = true } }) fmt.Fprintln(w, "\tbreakTest = []string{") parse("auxiliary/WordBreakTest.txt", func(p *ucd.Parser) { c := strings.Split(p.String(0), " ") const sep = '|' numCased := 0 test := "" for ; len(c) >= 2; c = c[2:] { if c[0] == "รท" && test != "" { test += string(sep) } i, err := strconv.ParseUint(c[1], 16, 32) r := rune(i) if err != nil { log.Fatalf("Invalid rune %q.", c[1]) } if r == sep { log.Fatalf("Separator %q not allowed in test data. Pick another one.", sep) } if cased[r] { numCased++ } test += string(r) } if numCased > 1 { fmt.Fprintf(w, "\t\t%q,\n", test) } }) fmt.Fprintln(w, "\t}") fmt.Fprintln(w, ")") gen.WriteGoFile("tables_test.go", "cases", w.Bytes()) } // These functions are just used for verification that their definition have not // changed in the Unicode Standard. func verifyCased(r rune) bool { return verifyLower(r) || verifyUpper(r) || unicode.IsTitle(r) } func verifyLower(r rune) bool { return unicode.IsLower(r) || unicode.Is(unicode.Other_Lowercase, r) } func verifyUpper(r rune) bool { return unicode.IsUpper(r) || unicode.Is(unicode.Other_Uppercase, r) } // verifyIgnore is an approximation of the Case_Ignorable property using the // core unicode package. It is used to reduce the size of the test data. func verifyIgnore(r rune) bool { props := []*unicode.RangeTable{ unicode.Mn, unicode.Me, unicode.Cf, unicode.Lm, unicode.Sk, } for _, p := range props { if unicode.Is(p, r) { return true } } return false } // printProperties prints tables of rune properties from the given UCD file. // A filter func f can be given to exclude certain values. A rune r will have // the indicated property if it is in the generated table or if f(r). func printProperties(w io.Writer, file, property string, f func(r rune) bool) int { verify := map[rune]bool{} n := 0 varNameParts := strings.Split(property, "_") varNameParts[0] = strings.ToLower(varNameParts[0]) fmt.Fprintf(w, "\t%s = map[rune]bool{\n", strings.Join(varNameParts, "")) parse(file, func(p *ucd.Parser) { if p.String(1) == property { r := p.Rune(0) verify[r] = true if !f(r) { n++ fmt.Fprintf(w, "\t\t0x%.4x: true,\n", r) } } }) fmt.Fprint(w, "\t}\n\n") // Verify that f is correct, that is, it represents a subset of the property. for r := rune(0); r <= lastRuneForTesting; r++ { if !verify[r] && f(r) { log.Fatalf("Incorrect filter func for property %q.", property) } } return n } // The newCaseTrie, sparseValues and sparseOffsets definitions below are // placeholders referred to by gen_trieval.go. The real definitions are // generated by this program and written to tables.go. func newCaseTrie(int) int { return 0 } var ( sparseValues [0]valueRange sparseOffsets [0]uint16 )