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)) && is_quote(ch))
373 if (prev == '#' && n < 0)
374 /* '#' is not a known mark, so assume it is a comment */
376 if (prev == '/' && ch == '/' && tk.txt.len == 1 && n < 0) {
377 close_token(state, &tk);
378 restore_unget_state(state);
381 if (prev == '/' && ch == '*' && tk.txt.len == 1 && n < 0) {
382 close_token(state, &tk);
383 restore_unget_state(state);
388 if (tk.num != TK_error) {
389 close_token(state, &tk);
393 If we don't find a known mark, we will check for strings and comments
394 before assuming that we have an unknown mark
403 if (ignored & (1<<TK_mark))
412 Strings start with one of single quote, double quote, or back quote
413 and continue until a matching character on the same line. Any of
414 these characters can be included in the list of known marks and then
415 they will not be used for identifying strings.
417 Immediately following the close quote, one or two ASCII letters may
418 appear. These are somewhat like the arbitrary letters allowed in
419 "Numbers" above. They can be used by the language in various ways.
421 If 3 identical quote characters appear in a row and are
422 followed by a newline, then this forms a multi-line string which
423 continues until an identical triple quote appears on a line preceded
424 only by whitespace and followed immediately by 0-2 ASCII letters and a newline.
426 Multi-line strings may not extend beyond the end of the `code_node` in
429 Normal strings and multi-line strings are encoded as two different
436 ###### internal functions
437 static int is_quote(wchar_t ch)
439 return ch == '\'' || ch == '"' || ch == '`';
442 #### Multi-line strings
444 The multi-line string is checked for first. If they are being
445 ignored, we fall through and treat a triple quote as an empty string
446 followed by the start of a new string.
449 if (tk.txt.len == 3 &&
450 !(ignored & (1 << TK_multi_string)) &&
451 is_quote(tk.txt.txt[0]) &&
452 memcmp(tk.txt.txt, tk.txt.txt+1, 2) == 0 &&
453 is_newline(tk.txt.txt[3])) {
455 wchar_t first = tk.txt.txt[0];
458 while (!at_eon(state) && qseen < 3) {
459 ch = get_char(state);
460 if (is_newline(ch)) {
463 } else if (at_sol && ch == first) {
465 } else if (ch != ' ' && ch != '\t') {
471 /* Hit end of node - error.
472 * unget so the newline is seen,
473 * but return rest of string as an error.
477 close_token(state, &tk);
481 /* 2 letters are allowed */
482 ch = get_char(state);
484 ch = get_char(state);
486 ch = get_char(state);
487 /* Now we must have a newline, but we don't return it
490 close_token(state, &tk);
491 tk.num = TK_multi_string;
497 #### Single-line strings
499 The sequence of marks collected may be more than a single-line
500 string, so we reset to the start and collect characters until
501 we find a close quote or a newline.
503 If `TK_string` is ignored, then quote characters will appear as `TK_mark`s.
506 if (tk.txt.len && is_quote(tk.txt.txt[0]) &&
507 !(ignored & (1<<TK_string))) {
508 wchar_t first = tk.txt.txt[0];
509 reset_token(state, &tk);
510 ch = get_char(state);
512 while (!at_eon(state) && !is_newline(ch)) {
513 ch = get_char(state);
518 if (is_newline(ch)) {
523 close_token(state, &tk);
529 Single line comments may start with '`//`' or '`#`' providing that these
530 are not known marks. They continue to the end of the line.
532 Block comments start with '`/*`' if this is not a known mark. They
533 continue to the first occurrence of '`*/`' and may not contain any
534 occurrence of '`/*`'.
536 Block comments can be wholly within one line or can continue over
537 multiple lines. The multi-line version should be followed immediately
538 by a newline. The Linux kernel contains over 285000 multi-line
539 comments are only 34 are followed by characters other than white space
540 (which should be removed) or a backslash (only needed in macros). So
541 it would not suffer from this rule.
543 These two comment types are reported as two separate token types, and
544 consequently can be ignored separately. When ignored a comment is
545 still parsed, but is discarded.
551 ###### internal functions
552 static int is_line_comment(struct text txt)
554 return (txt.len >= 1 && txt.txt[0] == '#') ||
555 (txt.len >= 2 && txt.txt[0] == '/' &&
559 static int is_block_comment(struct text txt)
561 return txt.len >= 2 && txt.txt[0] == '/' &&
565 #### Single line comments
567 A single-line comment continues up to, but not including the newline
572 if (is_line_comment(tk.txt)) {
573 while (!is_newline(ch) && !at_eon(state))
574 ch = get_char(state);
577 close_token(state, &tk);
578 tk.num = TK_line_comment;
579 if (ignored & (1 << TK_line_comment))
586 The token text collected so far could exceed the comment, so we need
589 If we find an embedded `/*` we reset to just before the '/' and report
590 an error. That way the next thing to be parsed will be the rest of
591 the comment. This requires a double unget, so we need to save/restore
592 the unget state (explained later).
596 if (is_block_comment(tk.txt)) {
599 reset_token(state, &tk);
602 save_unget_state(state);
603 ch = get_char(state);
605 while (!at_eon(state) &&
606 (prev != '/' || ch != '*') &&
607 (prev != '*' || ch != '/')) {
611 save_unget_state(state);
612 ch = get_char(state);
614 close_token(state, &tk);
620 /* embedded. Need to unget twice! */
621 restore_unget_state(state);
626 tk.num = TK_block_comment;
627 if (newlines && !(ignored & (1<<TK_newline))) {
628 /* next char must be newline */
629 ch = get_char(state);
634 if (tk.num == TK_error ||
635 !(ignored & (1 << TK_block_comment)))
640 ### Indents, Newlines, and White Space.
642 Normally white space is ignored. However newlines can be important as
643 can indents, which are either after a newline or at the start of a
644 node (detected by `at_son()`);
646 ###### exported functions
647 static inline int is_newline(wchar_t ch)
649 return ch == '\n' || ch == '\f' || ch == '\v';
653 if (ch <= ' ' && !is_newline(ch)
657 If a line starts with more white-space than the previous non-blank
658 line - or if the first non-blank line in the document starts with any
659 white-space - then an "IN" is reported at the start of the line.
661 Before the next non-blank line which starts with less white space, or
662 at the latest at the end of the document, a matching "OUT" token
663 is reported. There will always be an exact match between "IN" and
666 It is possible for "OUT" to be followed (almost) immediately by an
667 "IN". This happens if, for example, the indent of three consecutive
668 lines are 0, 8, 4 spaces. Before the second line we report an
669 "IN". Before the third line we must report an "OUT", as 4 is less
670 than 8, then also an Ident as 4 is greater than 0.
676 For the purpose of measuring the length of white space, a tab adds at
677 least one space, and rounds up to a multiple of 8.
679 ###### exported functions
680 static inline int indent_tab(int indent)
685 We need to track the current levels of indent. This requires some
686 sort of stack as indent levels are pushed on and popped off. In
687 practice this stack is unlikely to often exceed 5 so we will used a
688 fixed stack of 20 indent levels. More than this will be silently
693 int indent_sizes[20];
697 Newlines can optionally be reported. Newlines within a block comment
698 or a multi-line string are not reported separately, but each of these
699 must be followed immediately by a newline so these constructs cannot
700 hide the fact that a newline was present.
702 When indents are being reported, the Newline which would normally be
703 reported immediately before the "IN" is delayed until after the
704 matching "OUT". This makes an indented section act like a
705 continuation of the previous line to some extent.
707 A blank line would normally be reported simply as two consecutive Newline
708 tokens. However if the subsequent line is indented (and indents are being
709 reported) then the right thing to do is less obvious as Newlines should be
710 delayed - but how many Newlines?
712 The approach we will take is to report the extra Newlines immediately after
713 the IN token, so the blank line is treated as though it were an indented
719 If we find a newline or white space at the start of a block, we keep
720 collecting spaces, tabs, and newlines until we find some real text.
721 Then depending on the indent we generate some number of tokens. These
722 will be a sequence of "Newline OUT" pairs representing a decrease
723 in indent, then either a Newline or an IN depending on whether the
724 next line is indented, then zero or more Newlines representing all the
725 blank lines that have been skipped.
727 When a Newline leads to the next block of code there is a question of
728 whether the various Newline and OUT/IN tokens should appear to
729 pbelong to the earlier or later block. This is addressed by processing
730 the tokens in two stages based on the relative indent levels of the
731 two blocks (each block has a base indent to which the actual indents
734 Any "Newline OUT" pairs needed to reduce the current indent to the
735 maximum of the base indents of the old and new blocks are generated
736 against the old block. Then if the next block does not have an
737 increased indent, one more "Newline" is generated.
739 If further "Newline OUT" pairs are needed to get to the indent
740 level of the 'next' block, they are generated against that block,
741 though the first Newline is suppressed (it having already been
744 Finally the Newline or IN for the first line of the new block is
745 generated, unless the Newline needs to be suppressed because it
746 appeared at the end of the previous block.
748 This means that a block may start with an OUT or an IN, but
749 will only start with a Newline if it actually starts with a blank
752 We will need to represent in the `token_state` where in this sequence
753 of delayed tokens we are. As `state.col` records the target indent we
754 don't need to record how many OUTs or INs are needed. We do
755 need to record the number of blank lines, and which of Newline and
756 OUT is needed next in the initial sequence of pairs.
758 For this we store one more than the number of blank lines as
759 `delayed_lines` and a flag for `out_next`.
766 Generating these tokens involve two separate pieces of code.
768 Firstly we need to recognise white space and count the indents and
769 newlines. These are recorded in the above state fields.
771 Separately we need, on each call to `token_next`, we need to check if
772 there are some delayed tokens and if so we need to advance the state
773 information and return one token.
776 if (is_newline(ch) || (at_son(state) && ch <= ' ')) {
778 int was_son = at_son(state);
779 if (ignored & (1<<TK_in)) {
782 if (ignored & (1<<TK_newline))
785 close_token(state, &tk);
788 // Indents are needed, so check all white space.
789 while (ch <= ' ' && !at_eon(state)) {
792 ch = get_char(state);
796 if (state->node->next &&
797 state->node->next->indent > state->node->indent)
798 state->col = state->node->next->indent;
800 state->col = state->node->indent;
803 state->delayed_lines = newlines;
804 state->out_next = was_son;
805 state->check_indent = 1;
810 ###### delayed tokens
812 if (state->check_indent || state->delayed_lines) {
813 if (state->col < state->indent_sizes[state->indent_level]) {
814 if (!state->out_next &&
815 !(ignored & (1<<TK_newline))) {
820 state->indent_level -= 1;
825 if (state->col > state->indent_sizes[state->indent_level] &&
826 state->indent_level < sizeof(state->indent_sizes)-1) {
827 state->indent_level += 1;
828 state->indent_sizes[state->indent_level] = state->col;
829 state->delayed_lines -= 1;
833 state->check_indent = 0;
834 if (state->delayed_lines && !(ignored & (1<<TK_newline))) {
836 state->delayed_lines -= 1;
839 state->delayed_lines = 0;
845 After the last newline in the file has been processed, a special
846 end-of-file token will be returned. any further attempts to get more
847 tokens will continue to return the same end-of-file token.
857 state->check_indent = 1;
864 ### Unknown Marks, or errors.
866 We have now handled all the possible known mark-like tokens.
867 If the token we have is not empty and `TK_mark` is allowed,
868 we have an unknown mark, otherwise this must be an error.
871 /* one unknown character */
872 close_token(state, &tk);
876 ## Tools For The Task
878 You may have noticed that are few gaps we left in the above -
879 functions used without first defining them. Doing so above would have
882 ### Character by character
884 As we walk through the various `code_node`s we need to process whole
885 Unicode codepoints, and keep track of which line and column we are on.
886 We will assume for now that any printing character uses one column,
887 though that is not true in general.
889 As the text in a `code_node` may include an indent that identifies it as
890 being code, we need to be careful to strip that. The `code_node` has
891 a flag that tells us whether or not we need to strip.
897 struct code_node *node;
902 ###### internal functions
904 static void do_strip(struct token_state *state)
906 if (state->node->needs_strip) {
908 while (n && state->node->code.txt[state->offset] == ' ') {
912 while (n == 4 && state->node->code.txt[state->offset] == '\t') {
919 static wint_t get_char(struct token_state *state)
925 if (state->node == NULL)
927 if (state->node->code.len <= state->offset) {
929 state->node = state->node->next;
930 while (state->node && state->node->code.txt == NULL);
932 if (state->node == NULL)
935 state->line = state->node->line_no;
936 state->col = state->node->indent;
941 memset(&mbstate, 0, sizeof(mbstate));
943 n = mbrtowc(&next, state->node->code.txt + state->offset,
944 state->node->code.len - state->offset,
946 if (n == -2 || n == 0) {
947 /* Not enough bytes - not really possible */
949 state->offset = state->node->code.len;
950 } else if (n == -1) {
953 next = 0x7f; // an illegal character
959 } else if (is_newline(next)) {
961 state->col = state->node->indent;
963 } else if (next == '\t') {
964 state->col = indent_tab(state->col);
969 We will sometimes want to "unget" the last character as it needs to be
970 considered again as part of the next token. So we need to store a
971 'previous' version of all metadata.
978 ###### before get_char
979 state->prev_offset = state->offset;
980 state->prev_line = state->line;
981 state->prev_col = state->col;
983 ###### internal functions
985 static void unget_char(struct token_state *state)
988 state->offset = state->prev_offset;
989 state->line = state->prev_line;
990 state->col = state->prev_col;
994 We occasionally need a double-unget, particularly for numbers and
995 block comments. We don't impose this cost on all scanning, but
996 require those code sections that need it to call `save_unget_state`
997 before each `get_char`, and then `restore_unget_state` when a
998 double-unget is needed.
1005 ###### internal functions
1006 static void save_unget_state(struct token_state *state)
1008 state->prev_offset2 = state->prev_offset;
1009 state->prev_line2 = state->prev_line;
1010 state->prev_col2 = state->prev_col;
1013 static void restore_unget_state(struct token_state *state)
1015 state->prev_offset = state->prev_offset2;
1016 state->prev_line = state->prev_line2;
1017 state->prev_col = state->prev_col2;
1020 At the start of a token we don't want to be at the end of a code block
1021 if we can help it. To avoid this possibility, we 'get' and 'unget' a
1022 single character. This will move into the next non-empty code block
1023 and leave the current pointer at the start of it.
1025 This has to happen _after_ dealing with delayed tokens as some of them
1026 must appear in the previous node. When we do this, we need to reset
1027 the data in the token.
1029 ###### delayed tokens
1030 if (at_eon(state)) {
1033 tk.node = state->node;
1035 tk.txt.txt = state->node->code.txt + state->offset;
1036 tk.line = state->line;
1037 tk.col = state->col;
1043 The current token is initialized to line up with the first character
1044 that we 'get' for each token. When we have, or might have, a full
1045 token we can call `close_token` to set the `len` of the token
1046 appropriately. This can safely be called multiple times.
1048 Finally we occasionally (for single-line strings and block comments)
1049 need to reset to the beginning of the current token as we might have
1050 parsed too much already. For that there is `reset_token`.
1053 tk.node = state->node;
1055 tk.txt.txt = state->node->code.txt + state->offset;
1056 tk.line = state->line;
1057 tk.col = state->col;
1060 ###### internal functions
1062 static void close_token(struct token_state *state,
1065 tk->txt.len = (state->node->code.txt + state->offset)
1069 static void reset_token(struct token_state *state, struct token *tok)
1071 state->prev_line = tok->line;
1072 state->prev_col = tok->col;
1073 state->prev_offset = tok->txt.txt - state->node->code.txt;
1079 Tokens make not cross into the next `code_node`, and some tokens can
1080 include the newline at the and of a `code_node`, we must be able to
1081 easily check if we have reached the end. Equally we need to know if
1082 we are at the start of a node, as white space is treated a little
1085 ###### internal functions
1087 static int at_son(struct token_state *state)
1089 return state->offset == 0;
1092 static int at_eon(struct token_state *state)
1094 // at end-of-node ??
1095 return state->node == NULL ||
1096 state->offset >= state->node->code.len;
1099 ### Find a known word
1101 As the known-word list is sorted we can use a simple binary search.
1102 Following the pattern established in "mdcode", we will use a `struct
1103 text` with start and length to represent the code fragment we are
1106 ###### internal functions
1107 static int find_known(struct token_config *conf, struct text txt)
1110 int hi = conf->known_count;
1112 while (lo + 1 < hi) {
1113 int mid = (lo + hi) / 2;
1114 int cmp = strncmp(conf->words_marks[mid],
1116 if (cmp == 0 && conf->words_marks[mid][txt.len])
1123 if (strncmp(conf->words_marks[lo],
1124 txt.txt, txt.len) == 0
1125 && conf->words_marks[lo][txt.len] == 0)
1131 ### Bringing it all together
1133 Now we have all the bits there is just one section missing: combining
1134 all the token parsing code into one block.
1136 The handling of delayed tokens (Newlines, INs, OUTs) must come
1137 first before we try getting another character.
1139 Then we parse all the test, making sure that we check for known marks
1140 before strings and comments, but unknown marks after strings and comments.
1142 This block of code will either return a token, or will choose to
1143 ignore one, in which case it will `continue` around to the top of the
1149 ch = get_char(state);
1158 As well as getting tokens, we need to be able to create the
1159 `token_state` to start with, and discard it later.
1164 ###### main functions
1165 struct token_state *token_open(struct code_node *code, struct
1168 struct token_state *state = malloc(sizeof(*state));
1169 memset(state, 0, sizeof(*state));
1171 state->line = code->line_no;
1172 state->col = code->indent;
1177 void token_close(struct token_state *state)
1182 ###### exported functions
1183 struct token_state *token_open(struct code_node *code, struct
1184 token_config *conf);
1185 void token_close(struct token_state *state);
1189 Getting tokens is the main thing but it is also useful to be able to
1190 print out token information, particularly for tracing and testing.
1192 Known tokens are printed verbatim. Other tokens are printed as
1193 `type(content)` where content is truncated to a given number of characters.
1195 The function for printing a truncated string (`text_dump`) is also exported
1196 so that it can be used to tracing processed strings too.
1201 ###### exported functions
1202 void token_trace(FILE *f, struct token tok, int max);
1203 void text_dump(FILE *f, struct text t, int max);
1205 ###### main functions
1207 void text_dump(FILE *f, struct text txt, int max)
1214 for (i = 0; i < max; i++) {
1215 char c = txt.txt[i];
1216 if (c < ' ' || c > '~')
1217 fprintf(f, "\\x%02x", c & 0xff);
1221 fprintf(f, "%c", c);
1227 void token_trace(FILE *f, struct token tok, int max)
1229 static char *types[] = {
1230 [TK_ident] = "ident",
1232 [TK_number] = "number",
1233 [TK_string] = "string",
1234 [TK_multi_string] = "mstring",
1235 [TK_line_comment] = "lcomment",
1236 [TK_block_comment] = "bcomment",
1239 [TK_newline] = "newline",
1241 [TK_error] = "ERROR",
1245 default: /* known word or mark */
1246 fprintf(f, "%.*s", tok.txt.len, tok.txt.txt);
1252 /* No token text included */
1253 fprintf(f, "%s()", types[tok.num]);
1259 case TK_multi_string:
1260 case TK_line_comment:
1261 case TK_block_comment:
1263 fprintf(f, "%s(", types[tok.num]);
1264 text_dump(f, tok.txt, max);
1270 ### And there we have it
1272 We now have all the library functions defined for reading and printing
1273 tokens. Now we just need C files to store them, and a mk file to make them.
1275 ###### File: scanner.h
1277 ## exported functions
1279 ###### File: libscanner.c
1281 #include "scanner.h"
1283 ## internal functions
1286 ###### File: scanner.mk
1290 scanner.mk scanner.h libscanner.c : scanner.mdc
1293 libscanner.o : libscanner.c
1294 $(CC) $(CFLAGS) -c libscanner.c
1296 ## Processing numbers
1298 Converting a `TK_number` token to a numerical value is a slightly
1299 higher level task than lexical analysis, and slightly lower than
1300 grammar parsing, so put it here - as an index if you like.
1302 Importantly it will be used by the same testing rig that is used for
1303 testing the token scanner.
1305 The numeric value that we will convert all numbers into is the `mpq_t`
1306 from the GNU high precision number library "libgmp".
1308 ###### number includes
1312 Firstly we need to be able to parse a string of digits in a given base
1313 and possibly with a decimal marker. We store this in an `mpz_t`
1314 integer and report the number of digits after the decimal mark.
1316 On error we return zero and ensure that the 'mpz_t' has been freed, or
1317 had never been initialised.
1319 ###### number functions
1321 static int parse_digits(mpz_t num, struct text tok, int base,
1324 /* Accept digits up to 'base', ignore '_' and
1325 * ' ' if they appear between two legal digits,
1326 * and if `placesp` is not NULL, allow a single
1327 * '.' or ',' and report the number of digits
1329 * Return number of characters processed (p),
1330 * or 0 if something illegal was found.
1333 int decimal = -1; // digits after marker
1334 enum {Digit, Space, Other} prev = Other;
1337 for (p = 0; p < tok.len; p++) {
1339 char c = tok.txt[p];
1341 if (c == '_' || c == ' ') {
1347 if (c == '.' || c == ',') {
1350 if (!placesp || decimal >= 0)
1358 else if (isupper(c))
1360 else if (islower(c))
1371 mpz_mul_ui(num, num, base);
1375 mpz_add_ui(num, num, dig);
1394 ###### number includes
1397 To parse a full number we need to consider the optional base, the
1398 mantissa, and the optional exponent. We will treat these one at a
1401 The base is indicated by a letter after a leading zero, which must be
1402 followed by a base letter or a period. The base also determines the
1403 character which will mark an exponent.
1411 if (tok.txt[0] == '0' && tok.len > 1) {
1413 switch(tok.txt[1]) {
1444 // another digit is not permitted
1448 // must be decimal marker or trailing
1449 // letter, which are OK;
1456 After the base is the mantissa, which may contain a decimal mark, so
1457 we need to record the number of places. We won't impose the number of
1458 places until we have the exponent as well.
1465 ###### parse mantissa
1467 d = parse_digits(mant, tok, base, &places);
1473 mpq_set_z(num, mant);
1476 After the mantissa number may come an exponent which may be positive
1477 or negative. We assume at this point that we have seen the exponent
1485 ###### parse exponent
1487 if (tok.txt[0] == '+') {
1490 } else if (tok.txt[0] == '-') {
1496 d = parse_digits(exp, tok, 10, NULL);
1501 if (!mpz_fits_slong_p(exp)) {
1506 lexp = mpz_get_si(exp) * esign;
1512 Now that we have the mantissa and the exponent we can multiply them
1513 together, also allowing for the number of digits after the decimal
1516 For base 10, we simply subtract the decimal places from the exponent.
1517 For the other bases, as the exponent is alway based on 2, even for
1518 octal and hex, we need a bit more detail.
1519 We then recover the sign from the exponent, as division is quite
1520 different from multiplication.
1522 ###### calc exponent
1541 Imposing the exponent on the number is also very different for base 10
1542 than for the others. For the binary shift `gmp` provides a simple
1543 function. For base 10 we use something like Russian Peasant
1546 ###### calc exponent
1550 mpq_set_ui(tens, 10, 1);
1554 mpq_mul(num, num, tens);
1556 mpq_div(num, num, tens);
1561 mpq_mul(tens, tens, tens);
1566 mpq_mul_2exp(num, num, lexp);
1568 mpq_div_2exp(num, num, lexp);
1571 Now we are ready to parse a number: the base, mantissa, and exponent.
1572 If all goes well we check for the possible trailing letters and
1573 return. Return value is 1 for success and 0 for failure.
1576 ###### number functions
1577 int number_parse(mpq_t num, char tail[3], struct text tok)
1584 if (tok.len > 1 && (tok.txt[0] == expc ||
1585 tok.txt[0] == toupper(expc))) {
1592 for (i = 0; i < 2; i++) {
1595 if (!isalpha(tok.txt[i]))
1597 tail[i] = tok.txt[i];
1607 Number parsing goes in `libnumber.c`
1609 ###### File: libnumber.c
1617 ###### File: number.h
1618 int number_parse(mpq_t num, char tail[3], struct text tok);
1620 ###### File: scanner.mk
1622 libnumber.o : libnumber.c
1623 $(CC) $(CFLAGS) -c libnumber.c
1625 ## Processing strings
1627 Both `TK_string` and `TK_multi_string` require post-processing which
1628 can be one of two types: literal or with escapes processed.
1629 Even literal processing is non-trivial as the file may contain indents
1630 which need to be stripped.
1632 Errors can only occur when processing escapes. Any unrecognised
1633 character following the escape character will cause an error.
1635 Processing escapes and striping indents can only make the string
1636 shorter, not longer, so we allocate a buffer which is the same size as
1637 the string and process into that.
1639 To request escape processing, we pass the character we want to use for
1640 quoting, usually '`\`'. To avoid escape processing we pass a zero.
1643 int string_parse(struct token *tok, char escape,
1644 struct text *str, char tail[3])
1647 struct text t = tok->txt;
1651 if (tok->num == TK_string) {
1656 str->txt = malloc(t.len);
1669 The tail of the string can be 0, 1, or 2 letters
1672 if (i >= 0 && isalpha(t.txt[i-1]))
1674 if (i >= 0 && isalpha(t.txt[i-1]))
1676 strncpy(tail, t.txt+i, t.len-i);
1685 Stripping the quote of a single-line string is trivial.
1686 The only part that is at all interesting is that quote character must
1690 if (t.txt[t.len-1] != quote)
1700 For a multi-line string we have a little more work to do. We need to
1701 remove 3 quotes, not 1, and need to count the indent of the close
1702 quote as it will need to be stripped from all lines.
1706 t.txt[1] != quote || t.txt[2] != quote ||
1707 !is_newline(t.txt[3]))
1712 if (i <= 0 || t.txt[i-1] != quote)
1715 if (i <= 0 || t.txt[i-1] != quote)
1718 if (i <= 0 || t.txt[i-1] != quote)
1722 while (i > 0 && !is_newline(t.txt[i-1]))
1726 if (t.txt[i] == ' ')
1728 if (t.txt[i] == '\t')
1729 indent = indent_tab(indent);
1738 Now we just take one byte at a time. trans-ASCII unicode won't look
1739 like anything we are interested in so it will just be copied byte by
1744 for (i = 0; i < t.len; i++) {
1758 } else if (i+1 >= t.len) {
1759 // escape and end of string
1767 str->len = cp - str->txt;
1775 Every time we find a start of line, we strip spaces and tabs until the
1776 required indent is found.
1779 while (i < t.len && skipped < indent) {
1784 skipped = indent_tab(skipped);
1793 *cp++ = '\n'; break;
1795 *cp++ = '\r'; break;
1797 *cp++ = '\t'; break;
1799 *cp++ = '\b'; break;
1801 *cp++ = quote; break;
1803 *cp++ = '\f'; break;
1805 *cp++ = '\v'; break;
1807 *cp++ = '\a'; break;
1812 // 3 digit octal number
1815 if (t.txt[i+1] < '0' || t.txt[i+1] > '7' ||
1816 t.txt[i+2] < '0' || t.txt[i+1] > '7')
1818 n = (t.txt[i ]-'0') * 64 +
1819 (t.txt[i+1]-'0') * 8 +
1820 (t.txt[i+2]-'0') * 1;
1826 n = take_hex(2, t.txt+i+1, t.len-i-1);
1834 // 4 or 8 hex digits for unicode
1835 n = take_hex(c == 'u'?4:8, t.txt+i+1, t.len-i-1);
1838 memset(&pstate, 0, sizeof(pstate));
1839 n = wcrtomb(cp, n, &pstate);
1843 i += c == 'u' ? 4 : 8;
1848 else if (is_newline(c))
1858 For `\x` `\u` and `\U` we need to collect a specific number of
1861 ###### string functions
1863 static long take_hex(int digits, char *cp, int l)
1875 else if (isupper(c))
1886 #### File: libstring.c
1888 String parsing goes in `libstring.c`
1897 #include "scanner.h"
1901 ###### File: string.h
1902 int string_parse(struct token *tok, char escape,
1903 struct text *str, char tail[3]);
1905 ###### File: scanner.mk
1907 libstring.o : libstring.c
1908 $(CC) $(CFLAGS) -c libstring.c
1913 As "untested code is buggy code" we need a program to easily test
1914 the scanner library. This will simply parse a given file and report
1915 the tokens one per line.
1917 ###### File: scanner.c
1923 #include <sys/mman.h>
1929 #include "scanner.h"
1934 static void pr_err(char *msg)
1937 fprintf(stderr, "%s\n", msg);
1940 int main(int argc, char *argv[])
1945 struct token_state *state;
1946 const char *known[] = {
1955 struct token_config conf = {
1958 .words_marks = known,
1959 .number_chars = "., _+-",
1960 .known_count = sizeof(known)/sizeof(known[0]),
1961 .ignored = (0 << TK_line_comment)
1962 |(0 << TK_block_comment),
1964 struct section *table, *s, *prev;
1965 setlocale(LC_ALL,"");
1967 fprintf(stderr, "Usage: scanner file\n");
1970 fd = open(argv[1], O_RDONLY);
1972 fprintf(stderr, "scanner: cannot open %s: %s\n",
1973 argv[1], strerror(errno));
1976 len = lseek(fd, 0, 2);
1977 file = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0);
1978 table = code_extract(file, file+len, pr_err);
1981 (code_free(s->code), prev = s, s = s->next, free(prev))) {
1982 printf("Tokenizing: %.*s\n", s->section.len,
1984 state = token_open(s->code, &conf);
1986 struct token tk = token_next(state);
1987 printf("%d:%d ", tk.line, tk.col);
1988 token_trace(stdout, tk, 20);
1989 if (tk.num == TK_number) {
1992 if (number_parse(num, tail,tk.txt)) {
1993 printf(" %s ", tail);
1994 mpq_out_str(stdout, 10, num);
1997 printf(" BAD NUMBER");
1999 if (tk.num == TK_string ||
2000 tk.num == TK_multi_string) {
2004 if (tk.txt.txt[0] == '`')
2006 if (string_parse(&tk, esc,
2008 printf(" %s ", tail);
2009 text_dump(stdout, str, 20);
2012 printf(" BAD STRING");
2015 if (tk.num == TK_error)
2017 if (tk.num == TK_eof)
2023 ###### File: scanner.mk
2024 scanner.c : scanner.mdc
2027 scanner : scanner.o scanner.h libscanner.o libmdcode.o mdcode.h
2028 $(CC) $(CFLAGS) -o scanner scanner.o libscanner.o \
2029 libmdcode.o libnumber.o libstring.o -licuuc -lgmp
2030 scanner.o : scanner.c
2031 $(CC) $(CFLAGS) -c scanner.c