/* ******************************************************************************* * * Copyright (C) 1999, International Business Machines * Corporation and others. All Rights Reserved. * ******************************************************************************* * file name: store.c * encoding: US-ASCII * tab size: 8 (not used) * indentation:4 * * created on: 1999dec11 * created by: Markus W. Scherer * * Store Unicode character properties efficiently for * random access. */ #include #include #include "unicode/utypes.h" #include "unicode/uchar.h" #include "cmemory.h" #include "cstring.h" #include "filestrm.h" #include "unicode/udata.h" #include "unewdata.h" #include "genprops.h" #define DO_DEBUG_OUT 0 /* Unicode character properties file format ------------------------------------ The file format prepared and written here contains several data structures that store indexes or data. Before the data contents described below, there are the headers required by the udata API for loading ICU data. Especially, a UDataInfo structure precedes the actual data. It contains platform properties values and the file format version. The following is a description of format version 1.0 . Data contents: The contents is a parsed, binary form of several Unicode character database files, most prominently UnicodeData.txt. Any Unicode code point from 0 to 0x10ffff can be looked up to get the properties, if any, for that code point. This means that the input to the lookup are 21-bit unsigned integers, with not all of the 21-bit range used. It is assumed that client code keeps a uint16_t pointer to the beginning of the data: const uint16 *p16; Some indexes assume 32-bit units; although client code should only cast the above pointer to (const uint32_t *), it is easier here to talk about the result of the indexing with the definition of another pointer variable for this: const uint32_t *p32=(const uint32_t *)p16; Formally, the file contains the following structures: A0 const uint16_t STAGE_2_BITS(=6); A1 const uint16_t STAGE_3_BITS(=4); (STAGE_1_BITS(=11) not stored, implicitly=21-(STAGE_2_BITS+STAGE_3_BITS)) A2 const uint16_t exceptionsIndex; -- 32-bit unit index A3 const uint16_t reservedIndex; A4 const uint16_t reservedIndex; A5 const uint16_t reservedIndex; A6 const uint16_t reservedIndex; A7 const uint16_t reservedIndex; S1 const uint16_t stage1[0x440]; -- 0x440=0x110000>>10 S2 const uint16_t stage2[variable size]; S3 const uint16_t stage3[variable size]; (possible 1*uint16_t for padding to 4-alignment) P const uint32_t props32[variable size]; E const uint32_t exceptions[variable size]; 3-stage lookup and properties: In order to condense the data for the 21-bit code space, several properties of the Unicode code assignment are exploited: - The code space is sparse. - There are several 10k of consecutive codes with the same properties. - Characters and scripts are allocated in groups of 16 code points. - Inside blocks for scripts the properties are often repetitive. - The 21-bit space is not fully used for Unicode. The three-stage lookup organizes code points in groups of 16 in stage 3. 64 such groups are grouped again, resulting in blocks of 64 indexes for a total of 1k code points in stage 2. The first stage is limited according to all code points being <0x110000. Each stage contains indexes to groups or blocks of the next stage in an n:1 manner, i.e., multiple entries of one stage may index the same group or block in the next one. In the second and third stages, groups of 64 or 16 may partially or completely overlap to save space with repetitive properties. In the properties table, only unique 32-bit words are stored to exploit non-adjacent overlapping. This is why the third stage does not directly contain the 32-bit properties words but only indexes to them. The indexes in each stage take the offset in the data of the next block into account to save additional arithmetic in the access. With a given Unicode code point uint32_t c; and 0<=c<0x110000, the lookup uses the three stage tables to arrive at an index into the props32[] table containing the character properties for c. For some characters, not all of the properties can be efficiently encoded using 32 bits. For them, the 32-bit word contains an index into the exceptions[] array. The first stage consumes the 11 most significant bits of the 21-bit code point and results in an index into the second stage: uint16_t i2=p16[8+c>>10]; The second stage consumes bits 9 to 4 of c and results in an index into the third stage: uint16_t i3=p16[i2+((c>>4)&0x3f)]; The third stage consumes bits 3 to 0 of c and results in a code point- specific value, which itself is only an index into the props32[] table: uint16_t i=p16[i3+(c&0xf)]; Note that the bit numbers and shifts actually depend on the STAGE_2/3_BITS in p16[0..1]. There is finally the 32-bit encoded set of properties for c: uint32_t props=p32[i]; For some characters, this contains an index into the exceptions array: if(props&EXCEPTION_BIT)) { uint16_t e=(uint16_t)(props>>VALUE_SHIFT); ... } The exception values are a variable number of uint32_t starting at const uint32_t *pe=p32+exceptionsIndex+e; The first uint32_t there contains flags about what values actually follow it. Some of the exception values are UChar32 code points for the case mappings, others are numeric values etc. 32-bit properties sets: Each 32-bit properties word contains: 0.. 4 general category 5 has exception values 6..10 BiDi category 11 is mirrored 12..19 reserved 20..31 value according to bits 0..5: if(has exception) { exception index; } else switch(general category) { case Ll: delta to uppercase; -- same as titlecase case Lu: delta to lowercase; -- titlecase is same as c case Lt: delta to lowercase; -- uppercase is same as c case Mn: combining class; case N*: numeric value; default: if(is mirrored) { delta to mirror } else { 0 }; } Exception values: In the first uint32_t exception word for a code point, bits 31..24 reserved 23..16 combining class 15..0 flags that indicate which values follow: bit 0 has uppercase mapping 1 has lowercase mapping 2 has titlecase mapping 3 has numeric value (numerator) 4 has denominator value 5 has a mirror-image Unicode code point According to the flags in this word, one or more uint32_t words follow it in the sequence of the bit flags in the flags word; if a flag is not set, then the value is missing or 0: For the case mappings and the mirror-image Unicode code point, one uint32_t or UChar32 each is the code point. For the numeric/numerator value, an int32_t word contains the value directly, except for when there is no numerator but a denominator, then the numerator is 1. For the denominator value, a uint32_t word contains the value directly. Example: U+2160, ROMAN NUMERAL ONE, needs an exception because it has a lowercase mapping and a numeric value. Its exception values would be stored as 3 uint32_t words: - flags=0x0a (see above) with combining class 0 - lowercase mapping 0x2170 - numeric value=1 ----------------------------------------------------------------------------- */ /* UDataInfo cf. udata.h */ static UDataInfo dataInfo={ sizeof(UDataInfo), 0, U_IS_BIG_ENDIAN, U_CHARSET_FAMILY, U_SIZEOF_UCHAR, 0, 0x55, 0x50, 0x72, 0x6f, /* dataFormat="UPro" */ 1, 0, 0, 0, /* formatVersion */ 3, 0, 0, 0 /* dataVersion */ }; /* definitions and arrays for the 3-stage lookup */ enum { STAGE_2_BITS=6, STAGE_3_BITS=4, STAGE_1_BITS=21-(STAGE_2_BITS+STAGE_3_BITS), STAGE_2_SHIFT=STAGE_3_BITS, STAGE_1_SHIFT=(STAGE_2_SHIFT+STAGE_2_BITS), /* number of entries per sub-table in each stage */ STAGE_1_BLOCK=0x110000>>STAGE_1_SHIFT, STAGE_2_BLOCK=1<generalCategory==U_NON_SPACING_MARK; isNumber= genCategoryNames[p->generalCategory][0]=='N'; if(p->upperCase!=0) { /* verify that no numbers and no Mn have case mappings */ if(!(isMn || isNumber)) { value=(int32_t)p->code-(int32_t)p->upperCase; } else { x=EXCEPTION_BIT; } ++count; } if(p->lowerCase!=0) { /* verify that no numbers and no Mn have case mappings */ if(!(isMn || isNumber)) { value=(int32_t)p->lowerCase-(int32_t)p->code; } else { x=EXCEPTION_BIT; } ++count; } if(p->upperCase!=p->titleCase) { /* verify that no numbers and no Mn have case mappings */ if(!(isMn || isNumber)) { value=(int32_t)p->code-(int32_t)p->titleCase; } else { x=EXCEPTION_BIT; } ++count; } if(p->canonicalCombining>0) { /* verify that only Mn has a canonical combining class */ if(isMn) { value=p->canonicalCombining; } else { x=EXCEPTION_BIT; } ++count; } if(p->numericValue!=0) { /* verify that only numeric categories have numeric values */ if(isNumber) { value=p->numericValue; } else { x=EXCEPTION_BIT; } ++count; } if(p->denominator!=0) { /* verification for numeric category covered by the above */ value=p->denominator; ++count; } if(p->isMirrored) { if(p->mirrorMapping!=0) { value=(int32_t)p->mirrorMapping-(int32_t)p->code; } ++count; } /* handle exceptions */ if(count>1 || x!=0 || valuecode); } else if(count==1) { printf("*** code 0x%06x needs an exception because its value would be %ld\n", p->code, value); } else if(valuecode, value, MIN_VALUE,MAX_VALUE); } else { printf("*** code 0x%06x needs an exception because it has %u values\n", p->code, count); } } ++exceptionsCount; x=EXCEPTION_BIT; /* allocate and create exception values */ value=exceptionsTop; if(value>=4096) { fprintf(stderr, "genprops: out of exceptions memory at U+%06x. (%d exceeds allocated space)\n", p->code, value); exit(U_MEMORY_ALLOCATION_ERROR); } else { uint32_t first=(uint32_t)p->canonicalCombining<<16; uint16_t length=1; if(p->upperCase!=0) { first|=1; exceptions[value+length++]=p->upperCase; } if(p->lowerCase!=0) { first|=2; exceptions[value+length++]=p->lowerCase; } if(p->upperCase!=p->titleCase) { first|=4; exceptions[value+length++]=p->titleCase; } if(p->denominator==0) { if(p->numericValue!=0) { first|=8; exceptions[value+length++]=p->numericValue; } } else { if(p->numericValue!=1) { first|=8; exceptions[value+length++]=p->numericValue; } first|=0x10; exceptions[value+length++]=p->denominator; } if(p->isMirrored) { first|=0x20; exceptions[value+length++]=p->mirrorMapping; } exceptions[value]=first; exceptionsTop+=length; } } /* put together the 32-bit word of encoded properties */ x|= (uint32_t)p->generalCategory | (uint32_t)p->bidi<isMirrored<code, x, &count, &count, &count); if(beVerbose && p->code<=0x9f) { if(p->code==0) { printf("static uint32_t staticProps32Table[0xa0]={\n"); } if(x&EXCEPTION_BIT) { /* do something more intelligent if there is an exception */ printf(" /* 0x%02lx */ 0x%lx, /* has exception */\n", p->code, x&~EXCEPTION_BIT); } else { printf(" /* 0x%02lx */ 0x%lx,\n", p->code, x); } if(p->code==0x9f) { printf("};\n"); } } /* * "Higher-hanging fruit" (not implemented): * * For some sets of fields, there are fewer sets of values * than the product of the numbers of values per field. * This means that storing one single value for more than * one field and later looking up both field values in a table * saves space. * Examples: * - general category & BiDi * * There are only few common displacements between a code point * and its case mappings. Store deltas. Store codes for few * occuring deltas. */ } /* areas of same properties ------------------------------------------------- */ extern void repeatProps(void) { /* first and last code points in repetitive areas */ static const uint32_t areas[][2]={ { 0x3400, 0x4db5 }, /* CJK ext. A */ { 0x4e00, 0x9fa5 }, /* CJK */ { 0xac00, 0xd7a3 }, /* Hangul */ { 0xd800, 0xdb7f }, /* High Surrogates */ { 0xdb80, 0xdbff }, /* Private Use High Surrogates */ { 0xdc00, 0xdfff }, /* Low Surrogates */ { 0xe000, 0xf8ff }, /* Private Use */ { 0xf0000, 0x10ffff } /* Private Use */ }; /* * Set the repetitive properties for the big, known areas of all the same * character properties. Most of those will share the same stage 2 and 3 * tables. * * Assumptions: * - each area starts at a code point that is a multiple of 16 * - for each area, except the plane 15/16 private use one, * the first and last code points are defined in UnicodeData.txt * and thus already set * - there may be some properties already stored for some code points, * especially in the Private Use areas */ uint16_t i1, i2, i3, j3, i1Limit, i2Repeat, i3Repeat, area; uint32_t x, start, limit; /* set the properties for the plane 15/16 Private Use area if necessary */ if(getProps(0xf0000, &i1, &i2, &i3)==0) { setProps(0xf0000, getProps(0xe000, &i1, &i2, &i3), &i1, &i2, &i3); } /* fill in the repetitive properties */ for(area=0; area>STAGE_1_SHIFT); /* assume that i3 is the beginning of a stage 3 block (see assumptions above) */ /* is this stage 3 block suitable for setting it everywhere? (set i3Repeat) */ for(j3=1;;) { if(!(props[i3+j3]==0 || props[i3+j3]==x)) { /* this block contains different properties, we need a new one */ i3Repeat=allocProps(); break; } if(++j3==STAGE_3_BLOCK) { /* this block is good */ i3Repeat=i3; break; } } /* fill the repetitive block */ for(j3=0; j3>STAGE_2_SHIFT)&(STAGE_2_BLOCK-1))), i3Repeat, x); /* does this area end in an incomplete stage 3 block? */ repeatFromStage3(i2, (uint16_t)(limit&(STAGE_3_BLOCK-1)), x); return; } /* this stage 2 block will not be suitable for repetition */ i2Repeat=0; /* advance i1 to the first full block */ ++i1; } else { uint16_t j2; /* is this stage 2 block suitable for setting it everywhere? (set i2Repeat) */ for(j2=0;;) { if(!(stage2[i2+j2]==0 || stage2[i2+j2]==i3Repeat)) { /* this block contains different indexes, we will need a new one */ i2Repeat=0; break; } if(++j2==STAGE_2_BLOCK) { /* this block is good, set and fill it */ i2Repeat=i2; for(j2=0; j2>STAGE_2_SHIFT)&(STAGE_2_BLOCK-1))), i3Repeat, x); /* does this area end in an incomplete stage 3 block? */ repeatFromStage3(i2, (uint16_t)(limit&(STAGE_3_BLOCK-1)), x); } } } static uint16_t repeatFromStage2(uint16_t i2, uint16_t i2Limit, uint16_t i3Repeat, uint32_t x) { /* set a section of a stage 2 table and its properties to x */ uint16_t i3, j3; while(i20); } ++i2; } return i2; } static void repeatFromStage3(uint16_t i2, uint16_t j3, uint32_t x) { if(j3!=0) { /* fill in properties in a last, incomplete stage 3 block */ uint16_t i3=stage2[i2]; if(i3==0) { stage2[i2]=i3=allocProps(); } /* some properties may be set in this stage 3 block */ do { if(props[i3]==0) { props[i3]=x; } ++i3; } while(--j3>0); } } /* compacting --------------------------------------------------------------- */ extern void compactStage2(void) { uint16_t newTop=compactStage(stage2, stage2Top, STAGE_2_BLOCK, stage1, STAGE_1_BLOCK); /* we saved some space */ if(beVerbose) { printf("compactStage2() reduced stage2Top from %u to %u\n", stage2Top, newTop); } stage2Top=newTop; if(DO_DEBUG_OUT) { /* debug output */ uint16_t i1, i2, i3, i4; uint32_t c; for(c=0; c<0xffff; c+=307) { printf("properties(0x%06x)=0x%06x\n", c, getProps2(c, &i1, &i2, &i3, &i4)); } } } extern void compactStage3(void) { uint16_t newTop=compactStage(stage3, stage3Top, STAGE_3_BLOCK, stage2, stage2Top); /* we saved some space */ if(beVerbose) { printf("compactStage3() reduced stage3Top from %u to %u\n", stage3Top, newTop); } stage3Top=newTop; if(DO_DEBUG_OUT) { /* debug output */ uint16_t i1, i2, i3, i4; uint32_t c; for(c=0; c<0xffff; c+=307) { printf("properties(0x%06x)=0x%06x\n", c, getProps2(c, &i1, &i2, &i3, &i4)); } } } static uint16_t compactStage(uint16_t *stage, uint16_t stageTop, uint16_t blockSize, uint16_t *parent, uint16_t parentTop) { /* * This function is the common implementation for compacting * a stage table. * There are stageTop entries (indexes) in stage[]. * stageTop is a multiple of blockSize, and there are always blockSize stage[] entries * per parent stage entry which do not overlap - yet. * The first blockSize stage[] entries are always the empty ones. * We make the blocks overlap appropriately here and fill every blockSize-th entry in * map[] with the mapping from old to new properties indexes * in order to adjust the parent stage tables. * This simple algorithm does not find arbitrary overlaps, but only those * where the last i entries of the previous block and the first i of the * current one all have the same value. * This seems reasonable and yields linear performance. */ uint16_t i, start, prevEnd, newStart, x; map[0]=0; newStart=blockSize; for(start=newStart; start0; --i) { stage[newStart++]=stage[start++]; } } else if(newStart0; --i) { stage[newStart++]=stage[start++]; } } else /* no overlap && newStart==start */ { map[start]=start; newStart+=blockSize; start=newStart; } } /* now adjust the parent stage table */ for(i=0; i0;) { --i; map[i]=i; } /* do not reorder the first, empty entries */ qsort(map+STAGE_3_BLOCK, propsTop-STAGE_3_BLOCK, 2, compareProps); /* * Now invert the reordered table and compact it in the same step. * The result will be props32[] having only unique properties words * and stage3[] having indexes to them. */ newIndex=0; for(i=0; i>STAGE_1_SHIFT); *pI2=i2=stage1[i1]+(uint16_t)((c>>STAGE_2_SHIFT)&(STAGE_2_BLOCK-1)); *pI3=i3=stage2[i2]+(uint16_t)(c&(STAGE_3_BLOCK-1)); *pI4=i4=stage3[i3]; return props32[i4]; } /* get properties before compacting them */ static uint32_t getProps(uint32_t c, uint16_t *pI1, uint16_t *pI2, uint16_t *pI3) { uint16_t i1, i2, i3; *pI1=i1=(uint16_t)(c>>STAGE_1_SHIFT); *pI2=i2=stage1[i1]+(uint16_t)((c>>STAGE_2_SHIFT)&(STAGE_2_BLOCK-1)); *pI3=i3=stage2[i2]+(uint16_t)(c&(STAGE_3_BLOCK-1)); return props[i3]; } /* set properties before compacting them */ static void setProps(uint32_t c, uint32_t x, uint16_t *pI1, uint16_t *pI2, uint16_t *pI3) { uint16_t i1, i2, i3; *pI1=i1=(uint16_t)(c>>STAGE_1_SHIFT); i2=stage1[i1]; if(i2==0) { stage1[i1]=i2=allocStage2(); } *pI2=i2+=(uint16_t)((c>>STAGE_2_SHIFT)&(STAGE_2_BLOCK-1)); i3=stage2[i2]; if(i3==0) { stage2[i2]=i3=allocProps(); } *pI3=i3+=(uint16_t)(c&(STAGE_3_BLOCK-1)); props[i3]=x; } static uint16_t allocStage2(void) { uint16_t i=stage2Top; stage2Top+=STAGE_2_BLOCK; if(stage2Top>=MAX_STAGE_2_COUNT) { fprintf(stderr, "genprops: stage 2 overflow\n"); exit(U_MEMORY_ALLOCATION_ERROR); } return i; } static uint16_t allocProps(void) { uint16_t i=propsTop; propsTop+=STAGE_3_BLOCK; if(propsTop>=MAX_PROPS_COUNT) { fprintf(stderr, "genprops: properties overflow\n"); exit(U_MEMORY_ALLOCATION_ERROR); } return i; } static uint16_t addUChars(const UChar *s, uint16_t length) { uint16_t top=ucharsTop+length+1; UChar *p; if(top>=MAX_UCHAR_COUNT) { fprintf(stderr, "genprops: out of UChars memory\n"); exit(U_MEMORY_ALLOCATION_ERROR); } p=uchars+ucharsTop; uprv_memcpy(p, s, length); p[length]=0; ucharsTop=top; return (uint16_t)(p-uchars); } /* * Hey, Emacs, please set the following: * * Local Variables: * indent-tabs-mode: nil * End: * */