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 numbers 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 word 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 ensures that an unknown mark
328 immediately before a string is handled correctly.
330 If the first character of a comment marker (i.e. '/') is a known mark,
331 the above rules would suggest that the start of a comment would be
332 parsed as that mark, which is not what is wanted. So the introductory
333 sequences for a comment ("//" and "/*") are treated as
334 partially-known. They prevent the leading "/" from being a mark by
335 itself, but do not actually constitute a stand-alone mark.
337 If `TK_mark` is ignored, then unknown marks are returned as errors.
342 Known marks are included in the same list as the list of known words.
346 while (is_mark(ch, state->conf)) {
349 close_token(state, &tk);
350 n = find_known(state->conf, tk.txt);
352 tk.num = TK_reserved + n;
353 else if (tk.num != TK_error) {
354 /* found a longest-known-mark, still need to
357 if (tk.txt.len == 2 && tk.txt.txt[0] == '/' &&
358 (ch == '/' || ch == '*')) {
359 /* Yes, this is a comment, not a '/' */
360 restore_unget_state(state);
365 close_token(state, &tk);
369 save_unget_state(state);
370 ch = get_char(state);
371 if (!(ignored && (1<<TK_string)) && n < 0 &&is_quote(ch) && !is_quote(prev))
372 /* If strings are allowed, a quote (Which isn't a known mark)
373 * mustn't be treated as part of an unknown mark. It can be
374 * part of a multi-line srtings though.
377 if (prev == '#' && n < 0)
378 /* '#' is not a known mark, so assume it is a comment */
380 if (prev == '/' && ch == '/' && tk.txt.len == 1 && n < 0) {
381 close_token(state, &tk);
382 restore_unget_state(state);
385 if (prev == '/' && ch == '*' && tk.txt.len == 1 && n < 0) {
386 close_token(state, &tk);
387 restore_unget_state(state);
392 if (tk.num != TK_error) {
393 close_token(state, &tk);
397 If we don't find a known mark, we will check for strings and comments
398 before assuming that we have an unknown mark
407 if (ignored & (1<<TK_mark))
416 Strings start with one of single quote, double quote, or back quote
417 and continue until a matching character on the same line. Any of
418 these characters can be included in the list of known marks and then
419 they will not be used for identifying strings.
421 Immediately following the close quote, one or two ASCII letters may
422 appear. These are somewhat like the arbitrary letters allowed in
423 "Numbers" above. They can be used by the language in various ways.
425 If 3 identical quote characters appear in a row and are
426 followed by a newline, then this forms a multi-line string which
427 continues until an identical triple quote appears on a line preceded
428 only by whitespace and followed immediately by 0-2 ASCII letters and a newline.
430 Multi-line strings may not extend beyond the end of the `code_node` in
433 Normal strings and multi-line strings are encoded as two different
440 ###### internal functions
441 static int is_quote(wchar_t ch)
443 return ch == '\'' || ch == '"' || ch == '`'; // "
446 #### Multi-line strings
448 The multi-line string is checked for first. If they are being
449 ignored, we fall through and treat a triple quote as an empty string
450 followed by the start of a new string.
453 if (tk.txt.len == 3 &&
454 !(ignored & (1 << TK_multi_string)) &&
455 is_quote(tk.txt.txt[0]) &&
456 memcmp(tk.txt.txt, tk.txt.txt+1, 2) == 0 &&
457 is_newline(tk.txt.txt[3])) {
459 wchar_t first = tk.txt.txt[0];
462 while (!at_eon(state) && qseen < 3) {
463 ch = get_char(state);
464 if (is_newline(ch)) {
467 } else if (at_sol && ch == first) {
469 } else if (ch != ' ' && ch != '\t') {
475 /* Hit end of node - error.
476 * unget so the newline is seen,
477 * but return rest of string as an error.
481 close_token(state, &tk);
485 /* 2 letters are allowed */
486 ch = get_char(state);
488 ch = get_char(state);
490 ch = get_char(state);
491 /* Now we must have a newline, but we don't return it
494 close_token(state, &tk);
495 tk.num = TK_multi_string;
501 #### Single-line strings
503 The sequence of marks collected may be more than a single-line
504 string, so we reset to the start and collect characters until
505 we find a close quote or a newline.
507 If `TK_string` is ignored, then quote characters will appear as `TK_mark`s.
510 if (tk.txt.len && is_quote(tk.txt.txt[0]) &&
511 !(ignored & (1<<TK_string))) {
512 wchar_t first = tk.txt.txt[0];
513 reset_token(state, &tk);
514 ch = get_char(state);
516 while (!at_eon(state) && !is_newline(ch)) {
517 ch = get_char(state);
522 if (is_newline(ch)) {
527 while (!at_eon(state) && (ch = get_char(state)) &&
531 close_token(state, &tk);
537 Single line comments may start with '`//`' or '`#`' providing that these
538 are not known marks. They continue to the end of the line.
540 Block comments start with '`/*`' if this is not a known mark. They
541 continue to the first occurrence of '`*/`' and may not contain any
542 occurrence of '`/*`'.
544 Block comments can be wholly within one line or can continue over
545 multiple lines. The multi-line version should be followed immediately
546 by a newline. The Linux kernel contains over 285000 multi-line
547 comments are only 34 are followed by characters other than white space
548 (which should be removed) or a backslash (only needed in macros). So
549 it would not suffer from this rule.
551 These two comment types are reported as two separate token types, and
552 consequently can be ignored separately. When ignored a comment is
553 still parsed, but is discarded.
559 ###### internal functions
560 static int is_line_comment(struct text txt)
562 return (txt.len >= 1 && txt.txt[0] == '#') ||
563 (txt.len >= 2 && txt.txt[0] == '/' &&
567 static int is_block_comment(struct text txt)
569 return txt.len >= 2 && txt.txt[0] == '/' &&
573 #### Single line comments
575 A single-line comment continues up to, but not including the newline
580 if (is_line_comment(tk.txt)) {
581 while (!is_newline(ch) && !at_eon(state))
582 ch = get_char(state);
585 close_token(state, &tk);
586 tk.num = TK_line_comment;
587 if (ignored & (1 << TK_line_comment))
594 The token text collected so far could exceed the comment, so we need
597 If we find an embedded `/*` we reset to just before the '/' and report
598 an error. That way the next thing to be parsed will be the rest of
599 the comment. This requires a double unget, so we need to save/restore
600 the unget state (explained later).
604 if (is_block_comment(tk.txt)) {
607 reset_token(state, &tk);
610 save_unget_state(state);
611 ch = get_char(state);
613 while (!at_eon(state) &&
614 (prev != '/' || ch != '*') &&
615 (prev != '*' || ch != '/')) {
619 save_unget_state(state);
620 ch = get_char(state);
622 close_token(state, &tk);
628 /* embedded. Need to unget twice! */
629 restore_unget_state(state);
634 tk.num = TK_block_comment;
635 if (newlines && !(ignored & (1<<TK_newline))) {
636 /* next char must be newline */
637 ch = get_char(state);
642 if (tk.num == TK_error ||
643 !(ignored & (1 << TK_block_comment)))
648 ### Indents, Newlines, and White Space.
650 Normally white space is ignored. However newlines can be important as
651 can indents, which are either after a newline or at the start of a
652 node (detected by `at_son()`);
654 ###### exported functions
655 static inline int is_newline(wchar_t ch)
657 return ch == '\n' || ch == '\f' || ch == '\v';
661 if (ch <= ' ' && !is_newline(ch)
665 If a line starts with more white-space than the previous non-blank
666 line - or if the first non-blank line in the document starts with any
667 white-space - then an "IN" is reported at the start of the line.
669 Before the next non-blank line which starts with less white space, or
670 at the latest at the end of the document, a matching "OUT" token
671 is reported. There will always be an exact match between "IN" and
674 It is possible for "OUT" to be followed (almost) immediately by an
675 "IN". This happens if, for example, the indent of three consecutive
676 lines are 0, 8, 4 spaces. Before the second line we report an
677 "IN". Before the third line we must report an "OUT", as 4 is less
678 than 8, then also an Ident as 4 is greater than 0.
684 For the purpose of measuring the length of white space, a tab adds at
685 least one space, and rounds up to a multiple of 8.
687 ###### exported functions
688 static inline int indent_tab(int indent)
693 We need to track the current levels of indent. This requires some
694 sort of stack as indent levels are pushed on and popped off. In
695 practice this stack is unlikely to often exceed 5 so we will used a
696 fixed stack of 20 indent levels. More than this will be silently
701 int indent_sizes[20];
705 Newlines can optionally be reported. Newlines within a block comment
706 or a multi-line string are not reported separately, but each of these
707 must be followed immediately by a newline so these constructs cannot
708 hide the fact that a newline was present.
710 When indents are being reported, the Newline which would normally be
711 reported immediately before the "IN" is delayed until after the
712 matching "OUT". This makes an indented section act like a
713 continuation of the previous line to some extent.
715 A blank line would normally be reported simply as two consecutive Newline
716 tokens. However if the subsequent line is indented (and indents are being
717 reported) then the right thing to do is less obvious as Newlines should be
718 delayed - but how many Newlines?
720 The approach we will take is to report the extra Newlines immediately after
721 the IN token, so the blank line is treated as though it were an indented
727 If we find a newline or white space at the start of a block, we keep
728 collecting spaces, tabs, and newlines until we find some real text.
729 Then depending on the indent we generate some number of tokens. These
730 will be a sequence of "Newline OUT" pairs representing a decrease
731 in indent, then either a Newline or an IN depending on whether the
732 next line is indented, then zero or more Newlines representing all the
733 blank lines that have been skipped.
735 When a Newline leads to the next block of code there is a question of
736 whether the various Newline and OUT/IN tokens should appear to
737 pbelong to the earlier or later block. This is addressed by processing
738 the tokens in two stages based on the relative indent levels of the
739 two blocks (each block has a base indent to which the actual indents
742 Any "Newline OUT" pairs needed to reduce the current indent to the
743 maximum of the base indents of the old and new blocks are generated
744 against the old block. Then if the next block does not have an
745 increased indent, one more "Newline" is generated.
747 If further "Newline OUT" pairs are needed to get to the indent
748 level of the 'next' block, they are generated against that block,
749 though the first Newline is suppressed (it having already been
752 Finally the Newline or IN for the first line of the new block is
753 generated, unless the Newline needs to be suppressed because it
754 appeared at the end of the previous block.
756 This means that a block may start with an OUT or an IN, but
757 will only start with a Newline if it actually starts with a blank
760 We will need to represent in the `token_state` where in this sequence
761 of delayed tokens we are. As `state.col` records the target indent we
762 don't need to record how many OUTs or INs are needed. We do
763 need to record the number of blank lines, and which of Newline and
764 OUT is needed next in the initial sequence of pairs.
766 For this we store one more than the number of blank lines as
767 `delayed_lines` and a flag for `out_next`.
774 Generating these tokens involve two separate pieces of code.
776 Firstly we need to recognise white space and count the indents and
777 newlines. These are recorded in the above state fields.
779 Separately we need, on each call to `token_next`, we need to check if
780 there are some delayed tokens and if so we need to advance the state
781 information and return one token.
784 if (is_newline(ch) || (at_son(state) && ch <= ' ')) {
786 int was_son = at_son(state);
787 if (ignored & (1<<TK_in)) {
790 if (ignored & (1<<TK_newline))
793 close_token(state, &tk);
796 // Indents are needed, so check all white space.
797 while (ch <= ' ' && !at_eon(state)) {
800 ch = get_char(state);
804 if (state->node->next &&
805 state->node->next->indent > state->node->indent)
806 state->col = state->node->next->indent;
808 state->col = state->node->indent;
811 state->delayed_lines = newlines;
812 state->out_next = was_son;
813 state->check_indent = 1;
818 ###### delayed tokens
820 if (state->check_indent || state->delayed_lines) {
821 if (state->col < state->indent_sizes[state->indent_level]) {
822 if (!state->out_next &&
823 !(ignored & (1<<TK_newline))) {
828 state->indent_level -= 1;
833 if (state->col > state->indent_sizes[state->indent_level] &&
834 state->indent_level < sizeof(state->indent_sizes)-1) {
835 state->indent_level += 1;
836 state->indent_sizes[state->indent_level] = state->col;
837 state->delayed_lines -= 1;
841 state->check_indent = 0;
842 if (state->delayed_lines && !(ignored & (1<<TK_newline))) {
844 state->delayed_lines -= 1;
847 state->delayed_lines = 0;
853 After the last newline in the file has been processed, a special
854 end-of-file token will be returned. any further attempts to get more
855 tokens will continue to return the same end-of-file token.
865 state->check_indent = 1;
872 ### Unknown Marks, or errors.
874 We have now handled all the possible known mark-like tokens.
875 If the token we have is not empty and `TK_mark` is allowed,
876 we have an unknown mark, otherwise this must be an error.
879 /* one unknown character */
880 close_token(state, &tk);
884 ## Tools For The Task
886 You may have noticed that are few gaps we left in the above -
887 functions used without first defining them. Doing so above would have
890 ### Character by character
892 As we walk through the various `code_node`s we need to process whole
893 Unicode codepoints, and keep track of which line and column we are on.
894 We will assume for now that any printing character uses one column,
895 though that is not true in general.
897 As the text in a `code_node` may include an indent that identifies it as
898 being code, we need to be careful to strip that. The `code_node` has
899 a flag that tells us whether or not we need to strip.
905 struct code_node *node;
910 ###### internal functions
912 static int do_strip(struct token_state *state)
915 if (state->node->needs_strip) {
917 while (n && state->node->code.txt[state->offset] == ' ') {
922 while (n == 4 && state->node->code.txt[state->offset] == '\t') {
923 indent = indent_tab(indent);
931 static wint_t get_char(struct token_state *state)
937 if (state->node == NULL)
939 if (state->node->code.len <= state->offset) {
941 state->node = state->node->next;
942 while (state->node && state->node->code.txt == NULL);
944 if (state->node == NULL)
946 state->line = state->node->line_no;
947 state->col = do_strip(state);
952 memset(&mbstate, 0, sizeof(mbstate));
954 n = mbrtowc(&next, state->node->code.txt + state->offset,
955 state->node->code.len - state->offset,
957 if (n == -2 || n == 0) {
958 /* Not enough bytes - not really possible */
960 state->offset = state->node->code.len;
961 } else if (n == -1) {
964 next = 0x7f; // an illegal character
970 } else if (is_newline(next)) {
972 state->col = do_strip(state);
973 } else if (next == '\t') {
974 state->col = indent_tab(state->col);
979 We will sometimes want to "unget" the last character as it needs to be
980 considered again as part of the next token. So we need to store a
981 'previous' version of all metadata.
988 ###### before get_char
989 state->prev_offset = state->offset;
990 state->prev_line = state->line;
991 state->prev_col = state->col;
993 ###### internal functions
995 static void unget_char(struct token_state *state)
998 state->offset = state->prev_offset;
999 state->line = state->prev_line;
1000 state->col = state->prev_col;
1004 We occasionally need a double-unget, particularly for numbers and
1005 block comments. We don't impose this cost on all scanning, but
1006 require those code sections that need it to call `save_unget_state`
1007 before each `get_char`, and then `restore_unget_state` when a
1008 double-unget is needed.
1015 ###### internal functions
1016 static void save_unget_state(struct token_state *state)
1018 state->prev_offset2 = state->prev_offset;
1019 state->prev_line2 = state->prev_line;
1020 state->prev_col2 = state->prev_col;
1023 static void restore_unget_state(struct token_state *state)
1025 state->prev_offset = state->prev_offset2;
1026 state->prev_line = state->prev_line2;
1027 state->prev_col = state->prev_col2;
1030 At the start of a token we don't want to be at the end of a code block
1031 if we can help it. To avoid this possibility, we 'get' and 'unget' a
1032 single character. This will move into the next non-empty code block
1033 and leave the current pointer at the start of it.
1035 This has to happen _after_ dealing with delayed tokens as some of them
1036 must appear in the previous node. When we do this, we need to reset
1037 the data in the token.
1039 ###### delayed tokens
1040 if (at_eon(state)) {
1043 tk.node = state->node;
1045 tk.txt.txt = state->node->code.txt + state->offset;
1046 tk.line = state->line;
1047 tk.col = state->col;
1053 The current token is initialized to line up with the first character
1054 that we 'get' for each token. When we have, or might have, a full
1055 token we can call `close_token` to set the `len` of the token
1056 appropriately. This can safely be called multiple times.
1058 Finally we occasionally (for single-line strings and block comments)
1059 need to reset to the beginning of the current token as we might have
1060 parsed too much already. For that there is `reset_token`.
1063 tk.node = state->node;
1065 tk.txt.txt = state->node->code.txt + state->offset;
1066 tk.line = state->line;
1067 tk.col = state->col;
1070 ###### internal functions
1072 static void close_token(struct token_state *state,
1075 if (state->node != tk->node)
1076 tk->txt.len = tk->node->code.len - (tk->txt.txt - tk->node->code.txt);
1078 tk->txt.len = (state->node->code.txt + state->offset)
1082 static void reset_token(struct token_state *state, struct token *tok)
1084 state->prev_line = tok->line;
1085 state->prev_col = tok->col;
1086 state->prev_offset = tok->txt.txt - state->node->code.txt;
1092 Tokens make not cross into the next `code_node`, and some tokens can
1093 include the newline at the and of a `code_node`, we must be able to
1094 easily check if we have reached the end. Equally we need to know if
1095 we are at the start of a node, as white space is treated a little
1098 ###### internal functions
1100 static int at_son(struct token_state *state)
1102 return state->offset == 0;
1105 static int at_eon(struct token_state *state)
1107 // at end-of-node ??
1108 return state->node == NULL ||
1109 state->offset >= state->node->code.len;
1112 ### Find a known word
1114 As the known-word list is sorted we can use a simple binary search.
1115 Following the pattern established in "mdcode", we will use a `struct
1116 text` with start and length to represent the code fragment we are
1119 ###### internal functions
1120 static int find_known(struct token_config *conf, struct text txt)
1123 int hi = conf->known_count;
1125 while (lo + 1 < hi) {
1126 int mid = (lo + hi) / 2;
1127 int cmp = strncmp(conf->words_marks[mid],
1129 if (cmp == 0 && conf->words_marks[mid][txt.len])
1136 if (strncmp(conf->words_marks[lo],
1137 txt.txt, txt.len) == 0
1138 && conf->words_marks[lo][txt.len] == 0)
1144 ### Bringing it all together
1146 Now we have all the bits there is just one section missing: combining
1147 all the token parsing code into one block.
1149 The handling of delayed tokens (Newlines, INs, OUTs) must come
1150 first before we try getting another character.
1152 Then we parse all the test, making sure that we check for known marks
1153 before strings and comments, but unknown marks after strings and comments.
1155 This block of code will either return a token, or will choose to
1156 ignore one, in which case it will `continue` around to the top of the
1162 ch = get_char(state);
1171 As well as getting tokens, we need to be able to create the
1172 `token_state` to start with, and discard it later.
1177 ###### main functions
1178 struct token_state *token_open(struct code_node *code, struct
1181 struct token_state *state = malloc(sizeof(*state));
1182 memset(state, 0, sizeof(*state));
1184 state->line = code->line_no;
1185 state->col = do_strip(state);
1189 void token_close(struct token_state *state)
1194 ###### exported functions
1195 struct token_state *token_open(struct code_node *code, struct
1196 token_config *conf);
1197 void token_close(struct token_state *state);
1201 Getting tokens is the main thing but it is also useful to be able to
1202 print out token information, particularly for tracing and testing.
1204 Known tokens are printed verbatim. Other tokens are printed as
1205 `type(content)` where content is truncated to a given number of characters.
1207 The function for printing a truncated string (`text_dump`) is also exported
1208 so that it can be used to tracing processed strings too.
1213 ###### exported functions
1214 void token_trace(FILE *f, struct token tok, int max);
1215 void text_dump(FILE *f, struct text t, int max);
1217 ###### main functions
1219 void text_dump(FILE *f, struct text txt, int max)
1226 for (i = 0; i < max; i++) {
1227 char c = txt.txt[i];
1228 if (c < ' ' || c > '~')
1229 fprintf(f, "\\x%02x", c & 0xff);
1233 fprintf(f, "%c", c);
1239 void token_trace(FILE *f, struct token tok, int max)
1241 static char *types[] = {
1242 [TK_ident] = "ident",
1244 [TK_number] = "number",
1245 [TK_string] = "string",
1246 [TK_multi_string] = "mstring",
1247 [TK_line_comment] = "lcomment",
1248 [TK_block_comment] = "bcomment",
1251 [TK_newline] = "newline",
1253 [TK_error] = "ERROR",
1257 default: /* known word or mark */
1258 fprintf(f, "%.*s", tok.txt.len, tok.txt.txt);
1264 /* No token text included */
1265 fprintf(f, "%s()", types[tok.num]);
1271 case TK_multi_string:
1272 case TK_line_comment:
1273 case TK_block_comment:
1275 fprintf(f, "%s(", types[tok.num]);
1276 text_dump(f, tok.txt, max);
1282 ### And there we have it
1284 We now have all the library functions defined for reading and printing
1285 tokens. Now we just need C files to store them, and a mk file to make them.
1287 ###### File: scanner.h
1289 ## exported functions
1291 ###### File: libscanner.c
1293 #include "scanner.h"
1295 ## internal functions
1298 ###### File: scanner.mk
1302 scanner.mk scanner.h libscanner.c : scanner.mdc
1305 libscanner.o : libscanner.c
1306 $(CC) $(CFLAGS) -c libscanner.c
1308 ## Processing numbers
1310 Converting a `TK_number` token to a numerical value is a slightly
1311 higher level task than lexical analysis, and slightly lower than
1312 grammar parsing, so put it here - as an index if you like.
1314 Importantly it will be used by the same testing rig that is used for
1315 testing the token scanner.
1317 The numeric value that we will convert all numbers into is the `mpq_t`
1318 from the GNU high precision number library "libgmp".
1320 ###### number includes
1324 Firstly we need to be able to parse a string of digits in a given base
1325 and possibly with a decimal marker. We store this in an `mpz_t`
1326 integer and report the number of digits after the decimal mark.
1328 On error we return zero and ensure that the 'mpz_t' has been freed, or
1329 had never been initialised.
1331 ###### number functions
1333 static int parse_digits(mpz_t num, struct text tok, int base,
1336 /* Accept digits up to 'base', ignore '_' and
1337 * ' ' if they appear between two legal digits,
1338 * and if `placesp` is not NULL, allow a single
1339 * '.' or ',' and report the number of digits
1341 * Return number of characters processed (p),
1342 * or 0 if something illegal was found.
1345 int decimal = -1; // digits after marker
1346 enum {Digit, Space, Other} prev = Other;
1349 for (p = 0; p < tok.len; p++) {
1351 char c = tok.txt[p];
1353 if (c == '_' || c == ' ') {
1359 if (c == '.' || c == ',') {
1362 if (!placesp || decimal >= 0)
1370 else if (isupper(c))
1372 else if (islower(c))
1383 mpz_mul_ui(num, num, base);
1387 mpz_add_ui(num, num, dig);
1406 ###### number includes
1409 To parse a full number we need to consider the optional base, the
1410 mantissa, and the optional exponent. We will treat these one at a
1413 The base is indicated by a letter after a leading zero, which must be
1414 followed by a base letter or a period. The base also determines the
1415 character which will mark an exponent.
1423 if (tok.txt[0] == '0' && tok.len > 1) {
1425 switch(tok.txt[1]) {
1456 // another digit is not permitted
1460 // must be decimal marker or trailing
1461 // letter, which are OK;
1468 After the base is the mantissa, which may contain a decimal mark, so
1469 we need to record the number of places. We won't impose the number of
1470 places until we have the exponent as well.
1477 ###### parse mantissa
1479 d = parse_digits(mant, tok, base, &places);
1485 mpq_set_z(num, mant);
1488 After the mantissa number may come an exponent which may be positive
1489 or negative. We assume at this point that we have seen the exponent
1497 ###### parse exponent
1499 if (tok.txt[0] == '+') {
1502 } else if (tok.txt[0] == '-') {
1508 d = parse_digits(exp, tok, 10, NULL);
1513 if (!mpz_fits_slong_p(exp)) {
1518 lexp = mpz_get_si(exp) * esign;
1524 Now that we have the mantissa and the exponent we can multiply them
1525 together, also allowing for the number of digits after the decimal
1528 For base 10, we simply subtract the decimal places from the exponent.
1529 For the other bases, as the exponent is alway based on 2, even for
1530 octal and hex, we need a bit more detail.
1531 We then recover the sign from the exponent, as division is quite
1532 different from multiplication.
1534 ###### calc exponent
1553 Imposing the exponent on the number is also very different for base 10
1554 than for the others. For the binary shift `gmp` provides a simple
1555 function. For base 10 we use something like Russian Peasant
1558 ###### calc exponent
1562 mpq_set_ui(tens, 10, 1);
1566 mpq_mul(num, num, tens);
1568 mpq_div(num, num, tens);
1573 mpq_mul(tens, tens, tens);
1578 mpq_mul_2exp(num, num, lexp);
1580 mpq_div_2exp(num, num, lexp);
1583 Now we are ready to parse a number: the base, mantissa, and exponent.
1584 If all goes well we check for the possible trailing letters and
1585 return. Return value is 1 for success and 0 for failure.
1588 ###### number functions
1589 int number_parse(mpq_t num, char tail[3], struct text tok)
1596 if (tok.len > 1 && (tok.txt[0] == expc ||
1597 tok.txt[0] == toupper(expc))) {
1604 for (i = 0; i < 2; i++) {
1607 if (!isalpha(tok.txt[i]))
1609 tail[i] = tok.txt[i];
1619 Number parsing goes in `libnumber.c`
1621 ###### File: libnumber.c
1629 ###### File: number.h
1630 int number_parse(mpq_t num, char tail[3], struct text tok);
1632 ###### File: scanner.mk
1634 libnumber.o : libnumber.c
1635 $(CC) $(CFLAGS) -c libnumber.c
1637 ## Processing strings
1639 Both `TK_string` and `TK_multi_string` require post-processing which
1640 can be one of two types: literal or with escapes processed.
1641 Even literal processing is non-trivial as the file may contain indents
1642 which need to be stripped.
1644 Errors can only occur when processing escapes. Any unrecognised
1645 character following the escape character will cause an error.
1647 Processing escapes and striping indents can only make the string
1648 shorter, not longer, so we allocate a buffer which is the same size as
1649 the string and process into that.
1651 To request escape processing, we pass the character we want to use for
1652 quoting, usually '`\`'. To avoid escape processing we pass a zero.
1655 int string_parse(struct token *tok, char escape,
1656 struct text *str, char tail[3])
1659 struct text t = tok->txt;
1663 if (tok->num == TK_string) {
1668 str->txt = malloc(t.len);
1681 The tail of the string can be 0, 1, or 2 letters
1684 if (i >= 0 && isalpha(t.txt[i-1]))
1686 if (i >= 0 && isalpha(t.txt[i-1]))
1688 strncpy(tail, t.txt+i, t.len-i);
1697 Stripping the quote of a single-line string is trivial.
1698 The only part that is at all interesting is that quote character must
1702 if (t.txt[t.len-1] != quote)
1712 For a multi-line string we have a little more work to do. We need to
1713 remove 3 quotes, not 1, and need to count the indent of the close
1714 quote as it will need to be stripped from all lines.
1718 t.txt[1] != quote || t.txt[2] != quote ||
1719 !is_newline(t.txt[3]))
1724 if (i <= 0 || t.txt[i-1] != quote)
1727 if (i <= 0 || t.txt[i-1] != quote)
1730 if (i <= 0 || t.txt[i-1] != quote)
1734 while (i > 0 && !is_newline(t.txt[i-1]))
1738 if (t.txt[i] == ' ')
1740 if (t.txt[i] == '\t')
1741 indent = indent_tab(indent);
1750 Now we just take one byte at a time. trans-ASCII unicode won't look
1751 like anything we are interested in so it will just be copied byte by
1756 for (i = 0; i < t.len; i++) {
1770 } else if (i+1 >= t.len) {
1771 // escape and end of string
1779 str->len = cp - str->txt;
1787 Every time we find a start of line, we strip spaces and tabs until the
1788 required indent is found.
1791 while (i < t.len && skipped < indent) {
1796 skipped = indent_tab(skipped);
1805 *cp++ = '\n'; break;
1807 *cp++ = '\r'; break;
1809 *cp++ = '\t'; break;
1811 *cp++ = '\b'; break;
1813 *cp++ = quote; break;
1815 *cp++ = '\f'; break;
1817 *cp++ = '\v'; break;
1819 *cp++ = '\a'; break;
1824 // 3 digit octal number
1827 if (t.txt[i+1] < '0' || t.txt[i+1] > '7' ||
1828 t.txt[i+2] < '0' || t.txt[i+1] > '7')
1830 n = (t.txt[i ]-'0') * 64 +
1831 (t.txt[i+1]-'0') * 8 +
1832 (t.txt[i+2]-'0') * 1;
1838 n = take_hex(2, t.txt+i+1, t.len-i-1);
1846 // 4 or 8 hex digits for unicode
1847 n = take_hex(c == 'u'?4:8, t.txt+i+1, t.len-i-1);
1850 memset(&pstate, 0, sizeof(pstate));
1851 n = wcrtomb(cp, n, &pstate);
1855 i += c == 'u' ? 4 : 8;
1860 else if (is_newline(c))
1870 For `\x` `\u` and `\U` we need to collect a specific number of
1873 ###### string functions
1875 static long take_hex(int digits, char *cp, int l)
1887 else if (isupper(c))
1898 #### File: libstring.c
1900 String parsing goes in `libstring.c`
1909 #include "scanner.h"
1913 ###### File: string.h
1914 int string_parse(struct token *tok, char escape,
1915 struct text *str, char tail[3]);
1917 ###### File: scanner.mk
1919 libstring.o : libstring.c
1920 $(CC) $(CFLAGS) -c libstring.c
1925 As "untested code is buggy code" we need a program to easily test
1926 the scanner library. This will simply parse a given file and report
1927 the tokens one per line.
1929 ###### File: scanner.c
1935 #include <sys/mman.h>
1942 #include "scanner.h"
1947 static void pr_err(char *msg)
1950 fprintf(stderr, "%s\n", msg);
1953 static int kcmp(const void *ap, const void *bp)
1955 char * const *a = ap;
1956 char * const *b = bp;
1957 return strcmp(*a, *b);
1960 int main(int argc, char *argv[])
1965 char *filename = NULL;
1966 struct token_state *state;
1967 const char *known[] = {
1976 struct token_config conf = {
1979 .words_marks = known,
1980 .number_chars = "., _+-",
1981 .known_count = sizeof(known)/sizeof(known[0]),
1984 static const struct option long_options[] = {
1985 { "word-start", 1, NULL, 'W'},
1986 { "word-cont", 1, NULL, 'w'},
1987 { "number-chars", 1, NULL, 'n'},
1988 { "ignore-numbers", 0, NULL, 'N'},
1989 { "ignore-ident", 0, NULL, 'I'},
1990 { "ignore-marks", 0, NULL, 'M'},
1991 { "ignore-strings", 0, NULL, 'S'},
1992 { "ignore-multi-strings",0, NULL, 'z'},
1993 { "ignore-line-comment",0, NULL, 'c'},
1994 { "ignore-newline", 0, NULL, 'l'},
1995 { "ignore-block-comment", 0, NULL, 'C'},
1996 { "ignore-indent", 0, NULL, 'i'},
1997 { "file", 1, NULL, 'f'},
1998 { NULL, 0, NULL, 0},
2000 static const char options[] = "W:w:n:NIMSzclCif:";
2002 struct section *table, *s, *prev;
2005 setlocale(LC_ALL,"");
2006 while ((opt = getopt_long(argc, argv, options, long_options, NULL))
2009 case 'W': conf.word_start = optarg; break;
2010 case 'w': conf.word_cont = optarg; break;
2011 case 'n': conf.number_chars = optarg; break;
2012 case 'N': conf.ignored |= 1 << TK_number; break;
2013 case 'I': conf.ignored |= 1 << TK_ident; break;
2014 case 'M': conf.ignored |= 1 << TK_mark; break;
2015 case 'S': conf.ignored |= 1 << TK_string; break;
2016 case 'z': conf.ignored |= 1 << TK_multi_string; break;
2017 case 'c': conf.ignored |= 1 << TK_line_comment; break;
2018 case 'C': conf.ignored |= 1 << TK_block_comment; break;
2019 case 'l': conf.ignored |= 1 << TK_newline; break;
2020 case 'i': conf.ignored |= 1 << TK_in; break;
2021 case 'f': filename = optarg; break;
2022 default: fprintf(stderr, "scanner: unknown option '%c'.\n",
2028 if (optind < argc) {
2029 const char **wm = calloc(argc - optind, sizeof(char*));
2031 for (i = optind; i < argc; i++)
2032 wm[i - optind] = argv[i];
2033 qsort(wm, argc-optind, sizeof(char*), kcmp);
2034 conf.words_marks = wm;
2035 conf.known_count = argc - optind;
2039 fd = open(filename, O_RDONLY);
2043 fprintf(stderr, "scanner: cannot open %s: %s\n",
2044 filename, strerror(errno));
2047 len = lseek(fd, 0, 2);
2049 fprintf(stderr,"scanner: %s is empty or not seekable\n",
2050 filename ?: "stdin");
2053 file = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0);
2054 table = code_extract(file, file+len, pr_err);
2057 (code_free(s->code), prev = s, s = s->next, free(prev))) {
2058 printf("Tokenizing: %.*s\n", s->section.len,
2060 state = token_open(s->code, &conf);
2062 struct token tk = token_next(state);
2063 printf("%d:%d ", tk.line, tk.col);
2064 token_trace(stdout, tk, 20);
2065 if (tk.num == TK_number) {
2068 if (number_parse(num, tail,tk.txt)) {
2069 printf(" %s ", tail);
2070 mpq_out_str(stdout, 10, num);
2073 printf(" BAD NUMBER");
2075 if (tk.num == TK_string ||
2076 tk.num == TK_multi_string) {
2080 if (tk.txt.txt[0] == '`')
2082 if (string_parse(&tk, esc,
2084 printf(" %s ", tail);
2085 text_dump(stdout, str, 20);
2088 printf(" BAD STRING");
2091 if (tk.num == TK_error)
2093 if (tk.num == TK_eof)
2098 if (conf.words_marks != known)
2099 free(conf.words_marks);
2102 ###### File: scanner.mk
2103 scanner.c : scanner.mdc
2106 scanner : scanner.o scanner.h libscanner.o libmdcode.o mdcode.h
2107 $(CC) $(CFLAGS) -o scanner scanner.o libscanner.o \
2108 libmdcode.o libnumber.o libstring.o -licuuc -lgmp
2109 scanner.o : scanner.c
2110 $(CC) $(CFLAGS) -c scanner.c