5 The main task of the lexical scanner is to convert a stream of
6 characters into a stream of tokens. The tokens are then typically
7 used by a parser to extract the syntactic structure.
9 The stream of characters are assumed to be in memory identified by a
10 linked list of blocks, such as provided by the "[mdcode][]" literate
11 program extractor. A single token may never cross a block boundary.
18 The text is assumed to be UTF-8 though some matching assumes the
19 ASCII subset. If the text provided does not conform to UTF-8 an error
20 will be reported and some number of bytes will be skipped.
25 #include <unicode/uchar.h>
27 Tokens are returned by successive calls to the main interface
28 function: `token_next()` which has a `state` structure to keep track
29 of where it is up to. Each token carries not just a numeric
30 identifier but also the code block, the line and character within that
31 block, and the actual start and length using the `struct text` from
37 struct code_node *node;
48 ###### exported functions
49 struct token token_next(struct token_state *state);
52 struct token token_next(struct token_state *state)
63 The `line` and `col` offsets are useful for reporting errors.
64 The `txt` provides the content when that is important.
66 ### Token types and configuration ##
68 The scanner is not completely general, yet not completely specified.
69 There are a fixed set of token types, though particular tokens within
70 those types can be distinguish via configuration.
72 Most token types may be explicitly ignored, as typically comments
73 would be. The exact consequence of ignoring each token type varies
78 int ignored; // bit set of ignored tokens.
79 ## token config parameters
83 struct token_config *conf;
85 ###### token_next init
86 int ignored = state->conf->ignored;
89 The different tokens are numbers, words, marks, strings, comments,
90 newlines, EOF, and indents, each of which is examined in detail below.
92 There are various cases where no token can be found in part of the
93 input. All of these will be reported as a `TK_error` token.
95 It is possible to declare a number of strings which form distinct
96 tokens (rather than being grouped as e.g. 'word'). These are given
97 token numbers from `TK_reserved` upwards.
108 Numbers are the messiest tokens to parse, primarily because they can
109 contain characters that also have meaning outside of number and,
110 particularly, immediately after numbers.
112 The obvious example is the '`-`' sign. It can come inside a number for
113 a negative exponent, or after a number as a subtraction operator. To
114 be sure we have parsed as best as possible we need to only allow the
115 '`-`' inside a number if it is after an exponent character. This can be
116 `e` or `p` (for hex exponents), but `e` can also be a hexadecimal
117 digit, so we don't allow '`-`' after just any `e`.
119 To make matters worse, our language designer has decided to experiment
120 with allowing commas to be used as the decimal indicator, and spaces
121 to be used to separate groups of digits in large numbers. Both of
122 these can reasonably be restricted to appear between two digits, so we
123 have to add that condition to our tests.
125 So we cannot just treat numbers as starting with a digit and being
126 followed by some set of characters. We need more structure than that.
130 - Numbers must start with a digit.
131 - If the first digit is zero, the next character must be a base
132 signifier (one of `xob`) or a decimal marker (`.` or `,`).
133 In the first case the first `p` or `P` may be followed by a sign.
134 - If the number doesn't start with `0` followed by one of `xob`, the
135 first `e` may be followed by a sign.
136 - Any digit or hex digit may be followed by a space or underscore
137 providing that the subsequence character is also a (hex) digit.
138 This rule will require an extra level of 'unget' to be
139 supported when handling characters.
140 - Otherwise any digits or ASCII letters are allowed. We do not at
141 this point check that the digits given are permitted by the base.
142 That will happen when the token is converted to a number.
144 To allow easy configuration, the various non alphanumeric characters
145 are only permitted if they are listed in a configuration parameter.
147 ###### token config parameters
150 Note that numbers may not start with a period, so `.75` is not a
151 number. This is not the norm, but is not unheard of. Excluding these
152 numbers simplifies the rule at very little cost.
157 If TK_number is ignored, digits will result in an error unless they
158 are declared to be a start character for words.
166 if (iswdigit(ch) && !(ignored & (1<<TK_number))) {
167 int prev_special = 0;
169 int decimal_mark = 0;
171 wchar_t ch2 = get_char(state);
172 if (strchr("xobXOB", ch2) != NULL)
180 if (ch == 'e' || ch == 'E')
184 if (ch == 'p' || ch == 'P')
188 save_unget_state(state);
189 ch = get_char(state);
194 if (ch == '+' || ch == '-') {
199 if (ch == '.' || ch == ',') {
205 /* Don't allow that special char,
208 restore_unget_state(state);
211 if (strchr(state->conf->number_chars, ch)) {
215 /* non-number char */
218 /* We seem to have a "number" token */
220 close_token(state, &tk);
226 Words start with a "start" character followed by the longest
227 sequence of "continue" characters. The Unicode ID_START and
228 ID_CONTINUE sets are always permitted, but other ASCII characters
229 can be added to these sets.
231 ###### token config parameters
235 ###### internal functions
236 static int is_word_start(wchar_t ch, struct token_config *conf)
238 return iswalpha(ch) ||
239 strchr(conf->word_start, ch) != NULL ||
240 u_hasBinaryProperty(ch, UCHAR_ID_START);
243 static int is_word_continue(wchar_t ch, struct token_config *conf)
245 return iswalnum(ch) ||
246 strchr(conf->word_cont, ch) != NULL ||
247 u_hasBinaryProperty(ch, UCHAR_ID_CONTINUE);
250 Words can be either known or unknown. Known words are referred to as
251 "reserved words" and get a unique token number. Unknown words are
252 "identifiers" and are syntactically a single token.
257 A list of known words must be provided. This list is shared with the
258 "marks" which are described next. The list must be lexically sorted
259 and the length of the list must be given (`known_count`).
260 Tokens matching these known words are reported as the index of the
261 list added to `TK_reserved`.
263 If identifiers are ignored, then any work which is not listed as a
264 known word results in an error.
266 ###### token config parameters
267 const char **words_marks;
272 if (is_word_start(ch, state->conf)) {
274 /* A word: identifier or reserved */
276 ch = get_char(state);
277 while (is_word_continue(ch, state->conf));
279 close_token(state, &tk);
281 if (ignored & (1<<TK_ident))
283 n = find_known(state->conf, tk.txt);
285 tk.num = TK_reserved + n;
291 Marks are generally one or more punctuation marks joined together. It
292 would be nice to use the term "symbol" for these, but that causes
293 confusion in a subsequent discussion of the grammar, which has terminal
294 symbols and non-terminal symbols which are conceptually quite
295 different. So strings of punctuation characters will be marks.
297 A "mark" consists of ASCII characters that are not white space, are not
298 "start" characters for words, and are not digits.
299 These will collectively be called mark characters.
301 ###### internal functions
302 static int is_mark(wchar_t ch, struct token_config *conf)
307 strchr(conf->word_start, ch) == NULL;
310 As with words, there can be known and unknown marks, though the rules
311 are slightly different.
313 Two marks do not need to be separated by a non-mark characters. This
314 is different from words which do need to be separated by at least one
315 non-continue character.
317 The scanner will normally prefer longer sequences of mark characters,
318 but will more strongly prefer known marks over unknown marks. So if
319 it finds a known mark where adding one more character does not result
320 in a known mark, it will return that first known mark.
322 If no known mark is found we will test against strings and comments
323 below before giving up and assuming an unknown mark.
325 If an unknown mark contains a quote character or a comment marker, and
326 that token is not being ignored, then we terminate the unknown mark
327 before that quote or comment. This ensure that an unknown mark
328 immediately before a string is handled correctly.
330 If `TK_mark` is ignored, then unknown marks as returned as an error.
335 Known marks are included in the same list as the list of known words.
339 while (is_mark(ch, state->conf)) {
342 close_token(state, &tk);
343 n = find_known(state->conf, tk.txt);
345 tk.num = TK_reserved + n;
346 else if (tk.num != TK_error) {
347 /* found a longest-known-mark */
349 close_token(state, &tk);
353 save_unget_state(state);
354 ch = get_char(state);
355 if (!(ignored && (1<<TK_string)) && is_quote(ch))
359 if (prev == '/' && ch == '/' && tk.txt.len > 1) {
360 restore_unget_state(state);
363 if (prev == '/' && ch == '*' && tk.txt.len > 1) {
364 restore_unget_state(state);
369 if (tk.num != TK_error) {
370 close_token(state, &tk);
374 If we don't find a known mark, we will check for strings and comments
375 before assuming that we have an unknown mark
384 if (ignored & (1<<TK_mark))
393 Strings start with one of single quote, double quote, or back quote
394 and continue until a matching character on the same line. Any of
395 these characters can be included in the list of known marks and then
396 they will not be used for identifying strings.
398 Immediately following the close quote one or two ASCII letters may
399 appear. These are somewhat like the arbitrary letters allowed in
400 "Numbers" above. They can be used by the language in various ways.
402 If 3 identical quote characters appear in a row and are
403 followed by a newline, then this forms a multi-line string which
404 continues until an identical triple quote appears on a line preceded
405 only by whitespace and followed immediately by 0-2 ASCII letters and a newline.
407 Multi-line strings may not extend beyond the end of the `code_node` in
410 Normal strings and multi-line strings are encoded as two different
417 ###### internal functions
418 static int is_quote(wchar_t ch)
420 return ch == '\'' || ch == '"' || ch == '`';
423 #### Multi-line strings
425 The multi-line string is checked for first. If they are being
426 ignored, we fall through and treat a triple quote as an empty string
427 followed by the start of a new string.
430 if (tk.txt.len == 3 &&
431 !(ignored & (1 << TK_multi_string)) &&
432 is_quote(tk.txt.txt[0]) &&
433 memcmp(tk.txt.txt, tk.txt.txt+1, 2) == 0 &&
434 is_newline(tk.txt.txt[3])) {
436 wchar_t first = tk.txt.txt[0];
439 while (!at_eon(state) && qseen < 3) {
440 ch = get_char(state);
441 if (is_newline(ch)) {
444 } else if (at_sol && ch == first) {
446 } else if (ch != ' ' && ch != '\t') {
452 /* Hit end of node - error.
453 * unget so the newline is seen,
454 * but return rest of string as an error.
458 close_token(state, &tk);
462 /* 2 letters are allowed */
463 ch = get_char(state);
465 ch = get_char(state);
467 ch = get_char(state);
468 /* Now we must have a newline, but we don't return it
471 close_token(state, &tk);
472 tk.num = TK_multi_string;
478 #### Single-line strings
480 The sequence of marks collected may be more than a single-line
481 string, so we reset to the start and collect characters until
482 we find a close quote or a newline.
484 If `TK_string` is ignored, then quote characters will appear as `TK_mark`s.
487 if (tk.txt.len && is_quote(tk.txt.txt[0]) &&
488 !(ignored & (1<<TK_string))) {
489 wchar_t first = tk.txt.txt[0];
490 reset_token(state, &tk);
491 ch = get_char(state);
493 while (!at_eon(state) && !is_newline(ch)) {
494 ch = get_char(state);
499 if (is_newline(ch)) {
504 close_token(state, &tk);
510 Single line comments may start with '`//`' or '`#`' providing that these
511 are not known marks. They continue to the end of the line.
513 Block comments start with '`/*`' if this is not a known mark. They
514 continue to the first occurrence of '`*/`' and may not contain any
515 occurrence of '`/*`'.
517 Block comments can be wholly within one line or can continue over
518 multiple lines. The multi-line version should be followed immediately
519 by a newline. The Linux kernel contains over 285000 multi-line
520 comments are only 34 are followed by characters other than white space
521 (which should be removed) or a backslash (only needed in macros). So
522 it would not suffer from this rule.
524 These two comment types are reported as two separate token types, and
525 consequently can be ignored separately. When ignored a comment is
526 still parsed, but is discarded.
532 ###### internal functions
533 static int is_line_comment(struct text txt)
535 return (txt.len >= 1 && txt.txt[0] == '#') ||
536 (txt.len >= 2 && txt.txt[0] == '/' &&
540 static int is_block_comment(struct text txt)
542 return txt.len >= 2 && txt.txt[0] == '/' &&
546 #### Single line comments
548 A single-line comment continues up to, but not including the newline
553 if (is_line_comment(tk.txt)) {
554 while (!is_newline(ch) && !at_eon(state))
555 ch = get_char(state);
558 close_token(state, &tk);
559 tk.num = TK_line_comment;
560 if (ignored & (1 << TK_line_comment))
567 The token text collected so far could exceed the comment, so we need
570 If we find an embedded `/*` we reset to just before the '/' and report
571 an error. That way the next thing to be parsed will be the rest of
572 the comment. This requires a double unget, so we need to save/restore
573 the unget state (explained later).
577 if (is_block_comment(tk.txt)) {
580 reset_token(state, &tk);
583 save_unget_state(state);
584 ch = get_char(state);
586 while (!at_eon(state) &&
587 (prev != '/' || ch != '*') &&
588 (prev != '*' || ch != '/')) {
592 save_unget_state(state);
593 ch = get_char(state);
595 close_token(state, &tk);
601 /* embedded. Need to unget twice! */
602 restore_unget_state(state);
607 tk.num = TK_block_comment;
608 if (newlines && !(ignored & (1<<TK_newline))) {
609 /* next char must be newline */
610 ch = get_char(state);
615 if (tk.num == TK_error ||
616 !(ignored & (1 << TK_block_comment)))
621 ### Indents, Newlines, and White Space.
623 Normally white space is ignored. However newlines can be important as
624 can indents, which are either after a newline or at the start of a
625 node (detected by `at_son()`);
627 ###### exported functions
628 static inline int is_newline(wchar_t ch)
630 return ch == '\n' || ch == '\f' || ch == '\v';
634 if (ch <= ' ' && !is_newline(ch)
638 If a line starts with more white-space than the previous non-blank
639 line - or if the first non-blank line in the document starts with any
640 white-space - then an "IN" is reported at the start of the line.
642 Before the next non-blank line which starts with less white space, or
643 at the latest at the end of the document, a matching "OUT" token
644 is reported. There will always be an exact match between "IN" and
647 It is possible for "OUT" to be followed (almost) immediately by an
648 "IN". This happens if, for example, the indent of three consecutive
649 lines are 0, 8, 4 spaces. Before the second line we report an
650 "IN". Before the third line we must report an "OUT", as 4 is less
651 than 8, then also an Ident as 4 is greater than 0.
657 For the purpose of measuring the length of white space, a tab adds at
658 least one space, and rounds up to a multiple of 8.
660 ###### exported functions
661 static inline int indent_tab(int indent)
666 We need to track the current levels of indent. This requires some
667 sort of stack as indent levels are pushed on and popped off. In
668 practice this stack is unlikely to often exceed 5 so we will used a
669 fixed stack of 20 indent levels. More than this will be silently
674 int indent_sizes[20];
678 Newlines can optionally be reported. Newlines within a block comment
679 or a multi-line string are not reported separately, but each of these
680 must be followed immediately by a newline so these constructs cannot
681 hide the fact that a newline was present.
683 When indents are being reported, the Newline which would normally be
684 reported immediately before the "IN" is delayed until after the
685 matching "OUT". This makes an indented section act like a
686 continuation of the previous line to some extent.
688 A blank line would normally be reported simply as two consecutive Newline
689 tokens. However if the subsequent line is indented (and indents are being
690 reported) then the right thing to do is less obvious as Newlines should be
691 delayed - but how many Newlines?
693 The approach we will take is to report the extra Newlines immediately after
694 the IN token, so the blank line is treated as though it were an indented
700 If we find a newline or white space at the start of a block, we keep
701 collecting spaces, tabs, and newlines until we find some real text.
702 Then depending on the indent we generate some number of tokens. These
703 will be a sequence of "Newline OUT" pairs representing a decrease
704 in indent, then either a Newline or an IN depending on whether the
705 next line is indented, then zero or more Newlines representing all the
706 blank lines that have been skipped.
708 When a Newline leads to the next block of code there is a question of
709 whether the various Newline and OUT/IN tokens should appear to
710 pbelong to the earlier or later block. This is addressed by processing
711 the tokens in two stages based on the relative indent levels of the
712 two blocks (each block has a base indent to which the actual indents
715 Any "Newline OUT" pairs needed to reduce the current indent to the
716 maximum of the base indents of the old and new blocks are generated
717 against the old block. Then if the next block does not have an
718 increased indent, one more "Newline" is generated.
720 If further "Newline OUT" pairs are needed to get to the indent
721 level of the 'next' block, they are generated against that block,
722 though the first Newline is suppressed (it having already been
725 Finally the Newline or IN for the first line of the new block is
726 generated, unless the Newline needs to be suppressed because it
727 appeared at the end of the previous block.
729 This means that a block may start with an OUT or an IN, but
730 will only start with a Newline if it actually starts with a blank
733 We will need to represent in the `token_state` where in this sequence
734 of delayed tokens we are. As `state.col` records the target indent we
735 don't need to record how many OUTs or INs are needed. We do
736 need to record the number of blank lines, and which of Newline and
737 OUT is needed next in the initial sequence of pairs.
739 For this we store one more than the number of blank lines as
740 `delayed_lines` and a flag for `out_next`.
747 Generating these tokens involve two separate pieces of code.
749 Firstly we need to recognise white space and count the indents and
750 newlines. These are recorded in the above state fields.
752 Separately we need, on each call to `token_next`, we need to check if
753 there are some delayed tokens and if so we need to advance the state
754 information and return one token.
757 if (is_newline(ch) || (at_son(state) && ch <= ' ')) {
759 int was_son = at_son(state);
760 if (ignored & (1<<TK_in)) {
763 if (ignored & (1<<TK_newline))
766 close_token(state, &tk);
769 // Indents are needed, so check all white space.
770 while (ch <= ' ' && !at_eon(state)) {
773 ch = get_char(state);
777 if (state->node->next &&
778 state->node->next->indent > state->node->indent)
779 state->col = state->node->next->indent;
781 state->col = state->node->indent;
784 state->delayed_lines = newlines;
785 state->out_next = was_son;
786 state->check_indent = 1;
791 ###### delayed tokens
793 if (state->check_indent || state->delayed_lines) {
794 if (state->col < state->indent_sizes[state->indent_level]) {
795 if (!state->out_next &&
796 !(ignored & (1<<TK_newline))) {
801 state->indent_level -= 1;
806 if (state->col > state->indent_sizes[state->indent_level] &&
807 state->indent_level < sizeof(state->indent_sizes)-1) {
808 state->indent_level += 1;
809 state->indent_sizes[state->indent_level] = state->col;
810 state->delayed_lines -= 1;
814 state->check_indent = 0;
815 if (state->delayed_lines && !(ignored & (1<<TK_newline))) {
817 state->delayed_lines -= 1;
820 state->delayed_lines = 0;
826 After the last newline in the file has been processed, a special
827 end-of-file token will be returned. any further attempts to get more
828 tokens will continue to return the same end-of-file token.
838 state->check_indent = 1;
845 ### Unknown Marks, or errors.
847 We have now handled all the possible known mark-like tokens.
848 If the token we have is not empty and `TK_mark` is allowed,
849 we have an unknown mark, otherwise this must be an error.
852 /* one unknown character */
853 close_token(state, &tk);
857 ## Tools For The Task
859 You may have noticed that are few gaps we left in the above -
860 functions used without first defining them. Doing so above would have
863 ### Character by character
865 As we walk through the various `code_node`s we need to process whole
866 Unicode codepoints, and keep track of which line and column we are on.
867 We will assume for now that any printing character uses one column,
868 though that is not true in general.
870 As the text in a `code_node` may include an indent that identifies it as
871 being code, we need to be careful to strip that. The `code_node` has
872 a flag that tells us whether or not we need to strip.
878 struct code_node *node;
883 ###### internal functions
885 static void do_strip(struct token_state *state)
887 if (state->node->needs_strip) {
889 while (n && state->node->code.txt[state->offset] == ' ') {
893 while (n == 4 && state->node->code.txt[state->offset] == '\t') {
900 static wint_t get_char(struct token_state *state)
906 if (state->node == NULL)
908 if (state->node->code.len <= state->offset) {
910 state->node = state->node->next;
911 while (state->node && state->node->code.txt == NULL);
913 if (state->node == NULL)
916 state->line = state->node->line_no;
917 state->col = state->node->indent;
922 memset(&mbstate, 0, sizeof(mbstate));
924 n = mbrtowc(&next, state->node->code.txt + state->offset,
925 state->node->code.len - state->offset,
927 if (n == -2 || n == 0) {
928 /* Not enough bytes - not really possible */
930 state->offset = state->node->code.len;
931 } else if (n == -1) {
934 next = 0x7f; // an illegal character
940 } else if (is_newline(next)) {
942 state->col = state->node->indent;
944 } else if (next == '\t') {
945 state->col = indent_tab(state->col);
950 We will sometimes want to "unget" the last character as it needs to be
951 considered again as part of the next token. So we need to store a
952 'previous' version of all metadata.
959 ###### before get_char
960 state->prev_offset = state->offset;
961 state->prev_line = state->line;
962 state->prev_col = state->col;
964 ###### internal functions
966 static void unget_char(struct token_state *state)
969 state->offset = state->prev_offset;
970 state->line = state->prev_line;
971 state->col = state->prev_col;
975 We occasionally need a double-unget, particularly for numbers and
976 block comments. We don't impose this cost on all scanning, but
977 require those code sections that need it to call `save_unget_state`
978 before each `get_char`, and then `restore_unget_state` when a
979 double-unget is needed.
986 ###### internal functions
987 static void save_unget_state(struct token_state *state)
989 state->prev_offset2 = state->prev_offset;
990 state->prev_line2 = state->prev_line;
991 state->prev_col2 = state->prev_col;
994 static void restore_unget_state(struct token_state *state)
996 state->prev_offset = state->prev_offset2;
997 state->prev_line = state->prev_line2;
998 state->prev_col = state->prev_col2;
1001 At the start of a token we don't want to be at the end of a code block
1002 if we can help it. To avoid this possibility, we 'get' and 'unget' a
1003 single character. This will move into the next non-empty code block
1004 and leave the current pointer at the start of it.
1006 This has to happen _after_ dealing with delayed tokens as some of them
1007 must appear in the previous node. When we do this, we need to reset
1008 the data in the token.
1010 ###### delayed tokens
1011 if (at_eon(state)) {
1014 tk.node = state->node;
1016 tk.txt.txt = state->node->code.txt + state->offset;
1017 tk.line = state->line;
1018 tk.col = state->col;
1024 The current token is initialized to line up with the first character
1025 that we 'get' for each token. When we have, or might have, a full
1026 token we can call `close_token` to set the `len` of the token
1027 appropriately. This can safely be called multiple times.
1029 Finally we occasionally (for single-line strings and block comments)
1030 need to reset to the beginning of the current token as we might have
1031 parsed too much already. For that there is `reset_token`.
1034 tk.node = state->node;
1036 tk.txt.txt = state->node->code.txt + state->offset;
1037 tk.line = state->line;
1038 tk.col = state->col;
1041 ###### internal functions
1043 static void close_token(struct token_state *state,
1046 tk->txt.len = (state->node->code.txt + state->offset)
1050 static void reset_token(struct token_state *state, struct token *tok)
1052 state->prev_line = tok->line;
1053 state->prev_col = tok->col;
1054 state->prev_offset = tok->txt.txt - state->node->code.txt;
1060 Tokens make not cross into the next `code_node`, and some tokens can
1061 include the newline at the and of a `code_node`, we must be able to
1062 easily check if we have reached the end. Equally we need to know if
1063 we are at the start of a node, as white space is treated a little
1066 ###### internal functions
1068 static int at_son(struct token_state *state)
1070 return state->offset == 0;
1073 static int at_eon(struct token_state *state)
1075 // at end-of-node ??
1076 return state->node == NULL ||
1077 state->offset >= state->node->code.len;
1080 ### Find a known word
1082 As the known-word list is sorted we can use a simple binary search.
1083 Following the pattern established in "mdcode", we will use a `struct
1084 text` with start and length to represent the code fragment we are
1087 ###### internal functions
1088 static int find_known(struct token_config *conf, struct text txt)
1091 int hi = conf->known_count;
1093 while (lo + 1 < hi) {
1094 int mid = (lo + hi) / 2;
1095 int cmp = strncmp(conf->words_marks[mid],
1097 if (cmp == 0 && conf->words_marks[mid][txt.len])
1104 if (strncmp(conf->words_marks[lo],
1105 txt.txt, txt.len) == 0
1106 && conf->words_marks[lo][txt.len] == 0)
1112 ### Bringing it all together
1114 Now we have all the bits there is just one section missing: combining
1115 all the token parsing code into one block.
1117 The handling of delayed tokens (Newlines, INs, OUTs) must come
1118 first before we try getting another character.
1120 Then we parse all the test, making sure that we check for known marks
1121 before strings and comments, but unknown marks after strings and comments.
1123 This block of code will either return a token, or will choose to
1124 ignore one, in which case it will `continue` around to the top of the
1130 ch = get_char(state);
1139 As well as getting tokens, we need to be able to create the
1140 `token_state` to start with, and discard it later.
1145 ###### main functions
1146 struct token_state *token_open(struct code_node *code, struct
1149 struct token_state *state = malloc(sizeof(*state));
1150 memset(state, 0, sizeof(*state));
1152 state->line = code->line_no;
1153 state->col = code->indent;
1158 void token_close(struct token_state *state)
1163 ###### exported functions
1164 struct token_state *token_open(struct code_node *code, struct
1165 token_config *conf);
1166 void token_close(struct token_state *state);
1170 Getting tokens is the main thing but it is also useful to be able to
1171 print out token information, particularly for tracing and testing.
1173 Known tokens are printed verbatim. Other tokens are printed as
1174 `type(content)` where content is truncated to a given number of characters.
1176 The function for printing a truncated string (`text_dump`) is also exported
1177 so that it can be used to tracing processed strings too.
1182 ###### exported functions
1183 void token_trace(FILE *f, struct token tok, int max);
1184 void text_dump(FILE *f, struct text t, int max);
1186 ###### main functions
1188 void text_dump(FILE *f, struct text txt, int max)
1195 for (i = 0; i < max; i++) {
1196 char c = txt.txt[i];
1197 if (c < ' ' || c > '~')
1198 fprintf(f, "\\x%02x", c & 0xff);
1202 fprintf(f, "%c", c);
1208 void token_trace(FILE *f, struct token tok, int max)
1210 static char *types[] = {
1211 [TK_ident] = "ident",
1213 [TK_number] = "number",
1214 [TK_string] = "string",
1215 [TK_multi_string] = "mstring",
1216 [TK_line_comment] = "lcomment",
1217 [TK_block_comment] = "bcomment",
1220 [TK_newline] = "newline",
1222 [TK_error] = "ERROR",
1226 default: /* known word or mark */
1227 fprintf(f, "%.*s", tok.txt.len, tok.txt.txt);
1233 /* No token text included */
1234 fprintf(f, "%s()", types[tok.num]);
1240 case TK_multi_string:
1241 case TK_line_comment:
1242 case TK_block_comment:
1244 fprintf(f, "%s(", types[tok.num]);
1245 text_dump(f, tok.txt, max);
1251 ### And there we have it
1253 We now have all the library functions defined for reading and printing
1254 tokens. Now we just need C files to store them, and a mk file to make them.
1256 ###### File: scanner.h
1258 ## exported functions
1260 ###### File: libscanner.c
1262 #include "scanner.h"
1264 ## internal functions
1267 ###### File: scanner.mk
1271 scanner.mk scanner.h libscanner.c : scanner.mdc
1274 libscanner.o : libscanner.c
1275 $(CC) $(CFLAGS) -c libscanner.c
1277 ## Processing numbers
1279 Converting a `TK_number` token to a numerical value is a slightly
1280 higher level task than lexical analysis, and slightly lower than
1281 grammar parsing, so put it here - as an index if you like.
1283 Importantly it will be used by the same testing rig that is used for
1284 testing the token scanner.
1286 The numeric value that we will convert all numbers into is the `mpq_t`
1287 from the GNU high precision number library "libgmp".
1289 ###### number includes
1293 Firstly we need to be able to parse a string of digits in a given base
1294 and possibly with a decimal marker. We store this in an `mpz_t`
1295 integer and report the number of digits after the decimal mark.
1297 On error we return zero and ensure that the 'mpz_t' has been freed, or
1298 had never been initialised.
1300 ###### number functions
1302 static int parse_digits(mpz_t num, struct text tok, int base,
1305 /* Accept digits up to 'base', ignore '_' and
1306 * ' ' if they appear between two legal digits,
1307 * and if `placesp` is not NULL, allow a single
1308 * '.' or ',' and report the number of digits
1310 * Return number of characters processed (p),
1311 * or 0 if something illegal was found.
1314 int decimal = -1; // digits after marker
1315 enum {Digit, Space, Other} prev = Other;
1318 for (p = 0; p < tok.len; p++) {
1320 char c = tok.txt[p];
1322 if (c == '_' || c == ' ') {
1328 if (c == '.' || c == ',') {
1331 if (!placesp || decimal >= 0)
1339 else if (isupper(c))
1341 else if (islower(c))
1352 mpz_mul_ui(num, num, base);
1356 mpz_add_ui(num, num, dig);
1375 ###### number includes
1378 To parse a full number we need to consider the optional base, the
1379 mantissa, and the optional exponent. We will treat these one at a
1382 The base is indicated by a letter after a leading zero, which must be
1383 followed by a base letter or a period. The base also determines the
1384 character which will mark an exponent.
1392 if (tok.txt[0] == '0' && tok.len > 1) {
1394 switch(tok.txt[1]) {
1425 // another digit is not permitted
1429 // must be decimal marker or trailing
1430 // letter, which are OK;
1437 After the base is the mantissa, which may contain a decimal mark, so
1438 we need to record the number of places. We won't impose the number of
1439 places until we have the exponent as well.
1446 ###### parse mantissa
1448 d = parse_digits(mant, tok, base, &places);
1454 mpq_set_z(num, mant);
1457 After the mantissa number may come an exponent which may be positive
1458 or negative. We assume at this point that we have seen the exponent
1466 ###### parse exponent
1468 if (tok.txt[0] == '+') {
1471 } else if (tok.txt[0] == '-') {
1477 d = parse_digits(exp, tok, 10, NULL);
1482 if (!mpz_fits_slong_p(exp)) {
1487 lexp = mpz_get_si(exp) * esign;
1493 Now that we have the mantissa and the exponent we can multiply them
1494 together, also allowing for the number of digits after the decimal
1497 For base 10, we simply subtract the decimal places from the exponent.
1498 For the other bases, as the exponent is alway based on 2, even for
1499 octal and hex, we need a bit more detail.
1500 We then recover the sign from the exponent, as division is quite
1501 different from multiplication.
1503 ###### calc exponent
1522 Imposing the exponent on the number is also very different for base 10
1523 than for the others. For the binary shift `gmp` provides a simple
1524 function. For base 10 we use something like Russian Peasant
1527 ###### calc exponent
1531 mpq_set_ui(tens, 10, 1);
1535 mpq_mul(num, num, tens);
1537 mpq_div(num, num, tens);
1542 mpq_mul(tens, tens, tens);
1547 mpq_mul_2exp(num, num, lexp);
1549 mpq_div_2exp(num, num, lexp);
1552 Now we are ready to parse a number: the base, mantissa, and exponent.
1553 If all goes well we check for the possible trailing letters and
1554 return. Return value is 1 for success and 0 for failure.
1557 ###### number functions
1558 int number_parse(mpq_t num, char tail[3], struct text tok)
1565 if (tok.len > 1 && (tok.txt[0] == expc ||
1566 tok.txt[0] == toupper(expc))) {
1573 for (i = 0; i < 2; i++) {
1576 if (!isalpha(tok.txt[i]))
1578 tail[i] = tok.txt[i];
1588 Number parsing goes in `libnumber.c`
1590 ###### File: libnumber.c
1598 ###### File: number.h
1599 int number_parse(mpq_t num, char tail[3], struct text tok);
1601 ###### File: scanner.mk
1603 libnumber.o : libnumber.c
1604 $(CC) $(CFLAGS) -c libnumber.c
1606 ## Processing strings
1608 Both `TK_string` and `TK_multi_string` require post-processing which
1609 can be one of two types: literal or with escapes processed.
1610 Even literal processing is non-trivial as the file may contain indents
1611 which need to be stripped.
1613 Errors can only occur when processing escapes. Any unrecognised
1614 character following the escape character will cause an error.
1616 Processing escapes and striping indents can only make the string
1617 shorter, not longer, so we allocate a buffer which is the same size as
1618 the string and process into that.
1620 To request escape processing, we pass the character we want to use for
1621 quoting, usually '`\`'. To avoid escape processing we pass a zero.
1624 int string_parse(struct token *tok, char escape,
1625 struct text *str, char tail[3])
1628 struct text t = tok->txt;
1632 if (tok->num == TK_string) {
1637 str->txt = malloc(t.len);
1650 The tail of the string can be 0, 1, or 2 letters
1653 if (i >= 0 && isalpha(t.txt[i-1]))
1655 if (i >= 0 && isalpha(t.txt[i-1]))
1657 strncpy(tail, t.txt+i, t.len-i);
1666 Stripping the quote of a single-line string is trivial.
1667 The only part that is at all interesting is that quote character must
1671 if (t.txt[t.len-1] != quote)
1681 For a multi-line string we have a little more work to do. We need to
1682 remove 3 quotes, not 1, and need to count the indent of the close
1683 quote as it will need to be stripped from all lines.
1687 t.txt[1] != quote || t.txt[2] != quote ||
1688 !is_newline(t.txt[3]))
1693 if (i <= 0 || t.txt[i-1] != quote)
1696 if (i <= 0 || t.txt[i-1] != quote)
1699 if (i <= 0 || t.txt[i-1] != quote)
1703 while (i > 0 && !is_newline(t.txt[i-1]))
1707 if (t.txt[i] == ' ')
1709 if (t.txt[i] == '\t')
1710 indent = indent_tab(indent);
1719 Now we just take one byte at a time. trans-ASCII unicode won't look
1720 like anything we are interested in so it will just be copied byte by
1725 for (i = 0; i < t.len; i++) {
1739 } else if (i+1 >= t.len) {
1740 // escape and end of string
1748 str->len = cp - str->txt;
1756 Every time we find a start of line, we strip spaces and tabs until the
1757 required indent is found.
1760 while (i < t.len && skipped < indent) {
1765 skipped = indent_tab(c);
1774 *cp++ = '\n'; break;
1776 *cp++ = '\r'; break;
1778 *cp++ = '\t'; break;
1780 *cp++ = '\b'; break;
1782 *cp++ = quote; break;
1784 *cp++ = '\f'; break;
1786 *cp++ = '\v'; break;
1788 *cp++ = '\a'; break;
1793 // 3 digit octal number
1796 if (t.txt[i+1] < '0' || t.txt[i+1] > '7' ||
1797 t.txt[i+2] < '0' || t.txt[i+1] > '7')
1799 n = (t.txt[i ]-'0') * 64 +
1800 (t.txt[i+1]-'0') * 8 +
1801 (t.txt[i+2]-'0') * 1;
1807 n = take_hex(2, t.txt+i+1, t.len-i-1);
1815 // 4 or 8 hex digits for unicode
1816 n = take_hex(c == 'u'?4:8, t.txt+i+1, t.len-i-1);
1819 memset(&pstate, 0, sizeof(pstate));
1820 n = wcrtomb(cp, n, &pstate);
1824 i += c == 'u' ? 4 : 8;
1829 else if (is_newline(c))
1839 For `\x` `\u` and `\U` we need to collect a specific number of
1842 ###### string functions
1844 static long take_hex(int digits, char *cp, int l)
1856 else if (isupper(c))
1867 #### File: libstring.c
1869 String parsing goes in `libstring.c`
1878 #include "scanner.h"
1882 ###### File: string.h
1883 int string_parse(struct token *tok, char escape,
1884 struct text *str, char tail[3]);
1886 ###### File: scanner.mk
1888 libstring.o : libstring.c
1889 $(CC) $(CFLAGS) -c libstring.c
1894 As "untested code is buggy code" we need a program to easily test
1895 the scanner library. This will simply parse a given file and report
1896 the tokens one per line.
1898 ###### File: scanner.c
1904 #include <sys/mman.h>
1910 #include "scanner.h"
1915 static void pr_err(char *msg)
1918 fprintf(stderr, "%s\n", msg);
1921 int main(int argc, char *argv[])
1926 struct token_state *state;
1927 const char *known[] = {
1936 struct token_config conf = {
1939 .words_marks = known,
1940 .number_chars = "., _+-",
1941 .known_count = sizeof(known)/sizeof(known[0]),
1942 .ignored = (0 << TK_line_comment)
1943 |(0 << TK_block_comment),
1945 struct section *table, *s, *prev;
1946 setlocale(LC_ALL,"");
1948 fprintf(stderr, "Usage: scanner file\n");
1951 fd = open(argv[1], O_RDONLY);
1953 fprintf(stderr, "scanner: cannot open %s: %s\n",
1954 argv[1], strerror(errno));
1957 len = lseek(fd, 0, 2);
1958 file = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0);
1959 table = code_extract(file, file+len, pr_err);
1962 (code_free(s->code), prev = s, s = s->next, free(prev))) {
1963 printf("Tokenizing: %.*s\n", s->section.len,
1965 state = token_open(s->code, &conf);
1967 struct token tk = token_next(state);
1968 printf("%d:%d ", tk.line, tk.col);
1969 token_trace(stdout, tk, 20);
1970 if (tk.num == TK_number) {
1973 if (number_parse(num, tail,tk.txt)) {
1974 printf(" %s ", tail);
1975 mpq_out_str(stdout, 10, num);
1978 printf(" BAD NUMBER");
1980 if (tk.num == TK_string ||
1981 tk.num == TK_multi_string) {
1985 if (tk.txt.txt[0] == '`')
1987 if (string_parse(&tk, esc,
1989 printf(" %s ", tail);
1990 text_dump(stdout, str, 20);
1993 printf(" BAD STRING");
1996 if (tk.num == TK_error)
1998 if (tk.num == TK_eof)
2004 ###### File: scanner.mk
2005 scanner.c : scanner.mdc
2008 scanner : scanner.o scanner.h libscanner.o libmdcode.o mdcode.h
2009 $(CC) $(CFLAGS) -o scanner scanner.o libscanner.o \
2010 libmdcode.o libnumber.o libstring.o -licuuc -lgmp
2011 scanner.o : scanner.c
2012 $(CC) $(CFLAGS) -c scanner.c