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 an `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 ###### token config parameters
264 const char **words_marks;
269 if (is_word_start(ch, state->conf)) {
271 /* A word: identifier or reserved */
273 ch = get_char(state);
274 while (is_word_continue(ch, state->conf));
276 close_token(state, &tk);
278 if (ignored & (1<<TK_ident))
280 n = find_known(state->conf, tk.txt);
282 tk.num = TK_reserved + n;
288 Marks are generally one or more punctuation marks joined together. It
289 would be nice to use the term "symbol" for these, but that causes
290 confusion in a subsequent discussion of the grammar, which has terminal
291 symbols and non-terminal symbols which are conceptually quite
292 different. So strings of punctuation characters will be marks.
294 A "mark" consists of ASCII characters that are not white space, are not
295 "start" characters for words, and are not digits.
296 These will collectively be called mark characters.
298 ###### internal functions
299 static int is_mark(wchar_t ch, struct token_config *conf)
304 strchr(conf->word_start, ch) == NULL;
307 As with words, there can be known and unknown marks, though the rules
308 are slightly different.
310 Two marks do not need to be separated by a non-mark characters. This
311 is different from words which do need to be separated by at least one
312 non-continue character.
314 The scanner will normally prefer longer sequences of mark characters,
315 but will more strongly prefer known marks over unknown marks. So if
316 it finds a known mark where adding one more character does not result
317 in a known mark, it will return that first known mark.
319 If no known mark is found we will test against strings and comments
320 below before giving up and assuming an unknown mark.
322 If an unknown mark contains a quote character or a comment marker, and
323 that token is not being ignored, then we terminate the unknown mark
324 before that quote or comment. This ensure that an unknown mark
325 immediately before a string is handled correctly.
327 If `TK_mark` is ignored, then unknown marks as returned as an error.
332 Known marks are included in the same list as the list of known words.
336 while (is_mark(ch, state->conf)) {
339 close_token(state, &tk);
340 n = find_known(state->conf, tk.txt);
342 tk.num = TK_reserved + n;
343 else if (tk.num != TK_error) {
344 /* found a longest-known-mark */
346 close_token(state, &tk);
350 save_unget_state(state);
351 ch = get_char(state);
352 if (!(ignored && (1<<TK_string)) && is_quote(ch))
356 if (prev == '/' && ch == '/' && tk.txt.len > 1) {
357 restore_unget_state(state);
360 if (prev == '/' && ch == '*' && tk.txt.len > 1) {
361 restore_unget_state(state);
366 if (tk.num != TK_error) {
367 close_token(state, &tk);
371 If we don't find a known mark, we will check for strings and comments
372 before assuming that we have an unknown mark
381 if (ignored & (1<<TK_mark))
390 Strings start with one of single quote, double quote, or back quote
391 and continue until a matching character on the same line. Any of
392 these characters can be included in the list of known marks and then
393 they will not be used for identifying strings.
395 Immediately following the close quote one or two ASCII letters may
396 appear. These are somewhat like the arbitrary letters allowed in
397 "Numbers" above. They can be used by the language in various ways.
399 If 3 identical quote characters appear in a row and are
400 followed by a newline, then this forms a multi-line string which
401 continues until an identical triple quote appears on a line preceded
402 only by whitespace and followed immediately by 0-2 ASCII letters and a newline.
404 Multi-line strings may not extend beyond the end of the `code_node` in
407 Normal strings and multi-line strings are encoded as two different
414 ###### internal functions
415 static int is_quote(wchar_t ch)
417 return ch == '\'' || ch == '"' || ch == '`';
420 #### Multi-line strings
422 The multi-line string is checked for first. If they are being
423 ignored, we fall through and treat a triple quote as an empty string
424 followed by the start of a new string.
427 if (tk.txt.len == 3 &&
428 !(ignored & (1 << TK_multi_string)) &&
429 is_quote(tk.txt.txt[0]) &&
430 memcmp(tk.txt.txt, tk.txt.txt+1, 2) == 0 &&
431 is_newline(tk.txt.txt[3])) {
433 wchar_t first = tk.txt.txt[0];
436 while (!at_eon(state) && qseen < 3) {
437 ch = get_char(state);
438 if (is_newline(ch)) {
441 } else if (at_sol && ch == first) {
443 } else if (ch != ' ' && ch != '\t') {
449 /* Hit end of node - error.
450 * unget so the newline is seen,
451 * but return rest of string as an error.
455 close_token(state, &tk);
459 /* 2 letters are allowed */
460 ch = get_char(state);
462 ch = get_char(state);
464 ch = get_char(state);
465 /* Now we must have a newline, but we don't return it
468 close_token(state, &tk);
469 tk.num = TK_multi_string;
475 #### Single-line strings
477 The sequence of marks collected may be more than a single-line
478 string, so we reset to the start and collect characters until
479 we find a close quote or a newline.
481 If `TK_string` is ignored, then quote characters will appear as `TK_mark`s.
484 if (tk.txt.len && is_quote(tk.txt.txt[0]) &&
485 !(ignored & (1<<TK_string))) {
486 wchar_t first = tk.txt.txt[0];
487 reset_token(state, &tk);
488 ch = get_char(state);
490 while (!at_eon(state) && !is_newline(ch)) {
491 ch = get_char(state);
496 if (is_newline(ch)) {
501 close_token(state, &tk);
507 Single line comments may start with '`//`' or '`#`' providing that these
508 are not known marks. They continue to the end of the line.
510 Block comments start with '`/*`' if this is not a known mark. They
511 continue to the first occurrence of '`*/`' and may not contain any
512 occurrence of '`/*`'.
514 Block comments can be wholly within one line or can continue over
515 multiple lines. The multi-line version should be followed immediately
516 by a newline. The Linux kernel contains over 285000 multi-line
517 comments are only 34 are followed by characters other than white space
518 (which should be removed) or a backslash (only needed in macros). So
519 it would not suffer from this rule.
521 These two comment types are reported as two separate token types, and
522 consequently can be ignored separately. When ignored a comment is
523 still parsed, but is discarded.
529 ###### internal functions
530 static int is_line_comment(struct text txt)
532 return (txt.len >= 1 && txt.txt[0] == '#') ||
533 (txt.len >= 2 && txt.txt[0] == '/' &&
537 static int is_block_comment(struct text txt)
539 return txt.len >= 2 && txt.txt[0] == '/' &&
543 #### Single line comments
545 A single-line comment continues up to, but not including the newline
550 if (is_line_comment(tk.txt)) {
551 while (!is_newline(ch) && !at_eon(state))
552 ch = get_char(state);
555 close_token(state, &tk);
556 tk.num = TK_line_comment;
557 if (ignored & (1 << TK_line_comment))
564 The token text collected so far could exceed the comment, so we need
567 If we find an embedded `/*` we reset to just before the '/' and report
568 an error. That way the next thing to be parsed will be the rest of
569 the comment. This requires a double unget, so we need to save/restore
570 the unget state (explained later).
574 if (is_block_comment(tk.txt)) {
577 reset_token(state, &tk);
580 save_unget_state(state);
581 ch = get_char(state);
583 while (!at_eon(state) &&
584 (prev != '/' || ch != '*') &&
585 (prev != '*' || ch != '/')) {
589 save_unget_state(state);
590 ch = get_char(state);
592 close_token(state, &tk);
598 /* embedded. Need to unget twice! */
599 restore_unget_state(state);
604 tk.num = TK_block_comment;
605 if (newlines && !(ignored & (1<<TK_newline))) {
606 /* next char must be newline */
607 ch = get_char(state);
612 if (tk.num == TK_error ||
613 !(ignored & (1 << TK_block_comment)))
618 ### Indents, Newlines, and White Space.
620 Normally white space is ignored. However newlines can be important as
621 can indents, which are either after a newline or at the start of a
622 node (detected by `at_son()`);
624 ###### exported functions
625 static inline int is_newline(wchar_t ch)
627 return ch == '\n' || ch == '\f' || ch == '\v';
631 if (ch <= ' ' && !is_newline(ch)
635 If a line starts with more white-space than the previous non-blank
636 line - or if the first non-blank line in the document starts with any
637 white-space - then an "IN" is reported at the start of the line.
639 Before the next non-blank line which starts with less white space, or
640 at the latest at the end of the document, a matching "OUT" token
641 is reported. There will always be an exact match between "IN" and
644 It is possible for "OUT" to be followed (almost) immediately by an
645 "IN". This happens if, for example, the indent of three consecutive
646 lines are 0, 8, 4 spaces. Before the second line we report an
647 "IN". Before the third line we must report an "OUT", as 4 is less
648 than 8, then also an Ident as 4 is greater than 0.
654 For the purpose of measuring the length of white space, a tab adds at
655 least one space, and rounds up to a multiple of 8.
657 ###### exported functions
658 static inline int indent_tab(int indent)
663 We need to track the current levels of indent. This requires some
664 sort of stack as indent levels are pushed on and popped off. In
665 practice this stack is unlikely to often exceed 5 so we will used a
666 fixed stack of 20 indent levels. More than this will be silently
671 int indent_sizes[20];
675 Newlines can optionally be reported. Newlines within a block comment
676 or a multi-line string are not reported separately, but each of these
677 must be followed immediately by a newline so these constructs cannot
678 hide the fact that a newline was present.
680 When indents are being reported, the Newline which would normally be
681 reported immediately before the "IN" is delayed until after the
682 matching "OUT". This makes an indented section act like a
683 continuation of the previous line to some extent.
685 A blank line would normally be reported simply as two consecutive Newline
686 tokens. However if the subsequent line is indented (and indents are being
687 reported) then the right thing to do is less obvious as Newlines should be
688 delayed - but how many Newlines?
690 The approach we will take is to report the extra Newlines immediately after
691 the IN token, so the blank line is treated as though it were an indented
697 If we find a newline or white space at the start of a block, we keep
698 collecting spaces, tabs, and newlines until we find some real text.
699 Then depending on the indent we generate some number of tokens. These
700 will be a sequence of "Newline OUT" pairs representing a decrease
701 in indent, then either a Newline or an IN depending on whether the
702 next line is indented, then zero or more Newlines representing all the
703 blank lines that have been skipped.
705 When a Newline leads to the next block of code there is a question of
706 whether the various Newline and OUT/IN tokens should appear to
707 pbelong to the earlier or later block. This is addressed by processing
708 the tokens in two stages based on the relative indent levels of the
709 two blocks (each block has a base indent to which the actual indents
712 Any "Newline OUT" pairs needed to reduce the current indent to the
713 maximum of the base indents of the old and new blocks are generated
714 against the old block. Then if the next block does not have an
715 increased indent, one more "Newline" is generated.
717 If further "Newline OUT" pairs are needed to get to the indent
718 level of the 'next' block, they are generated against that block,
719 though the first Newline is suppressed (it having already been
722 Finally the Newline or IN for the first line of the new block is
723 generated, unless the Newline needs to be suppressed because it
724 appeared at the end of the previous block.
726 This means that a block may start with an OUT or an IN, but
727 will only start with a Newline if it actually starts with a blank
730 We will need to represent in the `token_state` where in this sequence
731 of delayed tokens we are. As `state.col` records the target indent we
732 don't need to record how many OUTs or INs are needed. We do
733 need to record the number of blank lines, and which of Newline and
734 OUT is needed next in the initial sequence of pairs.
736 For this we store one more than the number of blank lines as
737 `delayed_lines` and a flag for `out_next`.
744 Generating these tokens involve two separate pieces of code.
746 Firstly we need to recognise white space and count the indents and
747 newlines. These are recorded in the above state fields.
749 Separately we need, on each call to `token_next`, we need to check if
750 there are some delayed tokens and if so we need to advance the state
751 information and return one token.
754 if (is_newline(ch) || (at_son(state) && ch <= ' ')) {
756 int was_son = at_son(state);
757 if (ignored & (1<<TK_in)) {
760 if (ignored & (1<<TK_newline))
763 close_token(state, &tk);
766 // Indents are needed, so check all white space.
767 while (ch <= ' ' && !at_eon(state)) {
770 ch = get_char(state);
774 if (state->node->next &&
775 state->node->next->indent > state->node->indent)
776 state->col = state->node->next->indent;
778 state->col = state->node->indent;
781 state->delayed_lines = newlines;
782 state->out_next = was_son;
783 state->check_indent = 1;
788 ###### delayed tokens
790 if (state->check_indent || state->delayed_lines) {
791 if (state->col < state->indent_sizes[state->indent_level]) {
792 if (!state->out_next &&
793 !(ignored & (1<<TK_newline))) {
798 state->indent_level -= 1;
803 if (state->col > state->indent_sizes[state->indent_level] &&
804 state->indent_level < sizeof(state->indent_sizes)-1) {
805 state->indent_level += 1;
806 state->indent_sizes[state->indent_level] = state->col;
807 state->delayed_lines -= 1;
811 state->check_indent = 0;
812 if (state->delayed_lines && !(ignored & (1<<TK_newline))) {
814 state->delayed_lines -= 1;
817 state->delayed_lines = 0;
823 After the last newline in the file has been processed, a special
824 end-of-file token will be returned. any further attempts to get more
825 tokens will continue to return the same end-of-file token.
837 ### Unknown Marks, or errors.
839 We have now handled all the possible known mark-like tokens.
840 If the token we have is not empty and `TK_mark` is allowed,
841 we have an unknown mark, otherwise this must be an error.
844 /* one unknown character */
845 close_token(state, &tk);
849 ## Tools For The Task
851 You may have noticed that are few gaps we left in the above -
852 functions used without first defining them. Doing so above would have
855 ### Character by character
857 As we walk through the various `code_node`s we need to process whole
858 Unicode codepoints, and keep track of which line and column we are on.
859 We will assume for now that any printing character uses one column,
860 though that is not true in general.
862 As the text in a `code_node` may include an indent that identifies it as
863 being code, we need to be careful to strip that. The `code_node` has
864 a flag that tells us whether or not we need to strip.
870 struct code_node *node;
875 ###### internal functions
877 static void do_strip(struct token_state *state)
879 if (state->node->needs_strip) {
881 while (n && state->node->code.txt[state->offset] == ' ') {
885 while (n == 4 && state->node->code.txt[state->offset] == '\t') {
892 static wint_t get_char(struct token_state *state)
898 if (state->node == NULL)
900 if (state->node->code.len <= state->offset) {
902 state->node = state->node->next;
903 while (state->node && state->node->code.txt == NULL);
905 if (state->node == NULL)
908 state->line = state->node->line_no;
909 state->col = state->node->indent;
914 memset(&mbstate, 0, sizeof(mbstate));
916 n = mbrtowc(&next, state->node->code.txt + state->offset,
917 state->node->code.len - state->offset,
919 if (n == -2 || n == 0) {
920 /* Not enough bytes - not really possible */
922 state->offset = state->node->code.len;
923 } else if (n == -1) {
926 next = 0x7f; // an illegal character
932 } else if (is_newline(next)) {
934 state->col = state->node->indent;
936 } else if (next == '\t') {
937 state->col = indent_tab(state->col);
942 We will sometimes want to "unget" the last character as it needs to be
943 considered again as part of the next token. So we need to store a
944 'previous' version of all metadata.
951 ###### before get_char
952 state->prev_offset = state->offset;
953 state->prev_line = state->line;
954 state->prev_col = state->col;
956 ###### internal functions
958 static void unget_char(struct token_state *state)
961 state->offset = state->prev_offset;
962 state->line = state->prev_line;
963 state->col = state->prev_col;
967 We occasionally need a double-unget, particularly for numbers and
968 block comments. We don't impose this cost on all scanning, but
969 require those code sections that need it to call `save_unget_state`
970 before each `get_char`, and then `restore_unget_state` when a
971 double-unget is needed.
978 ###### internal functions
979 static void save_unget_state(struct token_state *state)
981 state->prev_offset2 = state->prev_offset;
982 state->prev_line2 = state->prev_line;
983 state->prev_col2 = state->prev_col;
986 static void restore_unget_state(struct token_state *state)
988 state->prev_offset = state->prev_offset2;
989 state->prev_line = state->prev_line2;
990 state->prev_col = state->prev_col2;
993 At the start of a token we don't want to be at the end of a code block
994 if we can help it. To avoid this possibility, we 'get' and 'unget' a
995 single character. This will move into the next non-empty code block
996 and leave the current pointer at the start of it.
998 This has to happen _after_ dealing with delayed tokens as some of them
999 must appear in the previous node. When we do this, we need to reset
1000 the data in the token.
1002 ###### delayed tokens
1003 if (at_eon(state)) {
1006 tk.node = state->node;
1008 tk.txt.txt = state->node->code.txt + state->offset;
1009 tk.line = state->line;
1010 tk.col = state->col;
1016 The current token is initialized to line up with the first character
1017 that we 'get' for each token. When we have, or might have, a full
1018 token we can call `close_token` to set the `len` of the token
1019 appropriately. This can safely be called multiple times.
1021 Finally we occasionally (for single-line strings and block comments)
1022 need to reset to the beginning of the current token as we might have
1023 parsed too much already. For that there is `reset_token`.
1026 tk.node = state->node;
1028 tk.txt.txt = state->node->code.txt + state->offset;
1029 tk.line = state->line;
1030 tk.col = state->col;
1033 ###### internal functions
1035 static void close_token(struct token_state *state,
1038 tk->txt.len = (state->node->code.txt + state->offset)
1042 static void reset_token(struct token_state *state, struct token *tok)
1044 state->prev_line = tok->line;
1045 state->prev_col = tok->col;
1046 state->prev_offset = tok->txt.txt - state->node->code.txt;
1052 Tokens make not cross into the next `code_node`, and some tokens can
1053 include the newline at the and of a `code_node`, we must be able to
1054 easily check if we have reached the end. Equally we need to know if
1055 we are at the start of a node, as white space is treated a little
1058 ###### internal functions
1060 static int at_son(struct token_state *state)
1062 return state->offset == 0;
1065 static int at_eon(struct token_state *state)
1067 // at end-of-node ??
1068 return state->node == NULL ||
1069 state->offset >= state->node->code.len;
1072 ### Find a known word
1074 As the known-word list is sorted we can use a simple binary search.
1075 Following the pattern established in "mdcode", we will use a `struct
1076 text` with start and length to represent the code fragment we are
1079 ###### internal functions
1080 static int find_known(struct token_config *conf, struct text txt)
1083 int hi = conf->known_count;
1085 while (lo + 1 < hi) {
1086 int mid = (lo + hi) / 2;
1087 int cmp = strncmp(conf->words_marks[mid],
1089 if (cmp == 0 && conf->words_marks[mid][txt.len])
1096 if (strncmp(conf->words_marks[lo],
1097 txt.txt, txt.len) == 0
1098 && conf->words_marks[lo][txt.len] == 0)
1104 ### Bringing it all together
1106 Now we have all the bits there is just one section missing: combining
1107 all the token parsing code into one block.
1109 The handling of delayed tokens (Newlines, INs, OUTs) must come
1110 first before we try getting another character.
1112 Then we parse all the test, making sure that we check for known marks
1113 before strings and comments, but unknown marks after strings and comments.
1115 This block of code will either return a token, or will choose to
1116 ignore one, in which case it will `continue` around to the top of the
1122 ch = get_char(state);
1131 As well as getting tokens, we need to be able to create the
1132 `token_state` to start with, and discard it later.
1137 ###### main functions
1138 struct token_state *token_open(struct code_node *code, struct
1141 struct token_state *state = malloc(sizeof(*state));
1142 memset(state, 0, sizeof(*state));
1144 state->line = code->line_no;
1145 state->col = code->indent;
1150 void token_close(struct token_state *state)
1155 ###### exported functions
1156 struct token_state *token_open(struct code_node *code, struct
1157 token_config *conf);
1158 void token_close(struct token_state *state);
1162 Getting tokens is the main thing but it is also useful to be able to
1163 print out token information, particularly for tracing and testing.
1165 Known tokens are printed verbatim. Other tokens are printed as
1166 `type(content)` where content is truncated to a given number of characters.
1168 The function for printing a truncated string (`text_dump`) is also exported
1169 so that it can be used to tracing processed strings too.
1174 ###### exported functions
1175 void token_trace(FILE *f, struct token tok, int max);
1176 void text_dump(FILE *f, struct text t, int max);
1178 ###### main functions
1180 void text_dump(FILE *f, struct text txt, int max)
1187 for (i = 0; i < max; i++) {
1188 char c = txt.txt[i];
1189 if (c < ' ' || c > '~')
1190 fprintf(f, "\\x%02x", c & 0xff);
1194 fprintf(f, "%c", c);
1200 void token_trace(FILE *f, struct token tok, int max)
1202 static char *types[] = {
1203 [TK_ident] = "ident",
1205 [TK_number] = "number",
1206 [TK_string] = "string",
1207 [TK_multi_string] = "mstring",
1208 [TK_line_comment] = "lcomment",
1209 [TK_block_comment] = "bcomment",
1212 [TK_newline] = "newline",
1214 [TK_error] = "ERROR",
1218 default: /* known word or mark */
1219 fprintf(f, "%.*s", tok.txt.len, tok.txt.txt);
1225 /* No token text included */
1226 fprintf(f, "%s()", types[tok.num]);
1232 case TK_multi_string:
1233 case TK_line_comment:
1234 case TK_block_comment:
1236 fprintf(f, "%s(", types[tok.num]);
1237 text_dump(f, tok.txt, max);
1243 ### And there we have it
1245 We now have all the library functions defined for reading and printing
1246 tokens. Now we just need C files to store them, and a mk file to make them.
1248 ###### File: scanner.h
1250 ## exported functions
1252 ###### File: libscanner.c
1254 #include "scanner.h"
1256 ## internal functions
1259 ###### File: scanner.mk
1263 scanner.mk scanner.h libscanner.c : scanner.mdc
1266 libscanner.o : libscanner.c
1267 $(CC) $(CFLAGS) -c libscanner.c
1269 ## Processing numbers
1271 Converting a `TK_number` token to a numerical value is a slightly
1272 higher level task than lexical analysis, and slightly lower than
1273 grammar parsing, so put it here - as an index if you like.
1275 Importantly it will be used by the same testing rig that is used for
1276 testing the token scanner.
1278 The numeric value that we will convert all numbers into is the `mpq_t`
1279 from the GNU high precision number library "libgmp".
1281 ###### number includes
1285 Firstly we need to be able to parse a string of digits in a given base
1286 and possibly with a decimal marker. We store this in an `mpz_t`
1287 integer and report the number of digits after the decimal mark.
1289 On error we return zero and ensure that the 'mpz_t' has been freed, or
1290 had never been initialised.
1292 ###### number functions
1294 static int parse_digits(mpz_t num, struct text tok, int base,
1297 /* Accept digits up to 'base', ignore '_' and
1298 * ' ' if they appear between two legal digits,
1299 * and if `placesp` is not NULL, allow a single
1300 * '.' or ',' and report the number of digits
1302 * Return number of characters processed (p),
1303 * or 0 if something illegal was found.
1306 int decimal = -1; // digits after marker
1307 enum {Digit, Space, Other} prev = Other;
1310 for (p = 0; p < tok.len; p++) {
1312 char c = tok.txt[p];
1314 if (c == '_' || c == ' ') {
1320 if (c == '.' || c == ',') {
1323 if (!placesp || decimal >= 0)
1331 else if (isupper(c))
1333 else if (islower(c))
1344 mpz_mul_ui(num, num, base);
1348 mpz_add_ui(num, num, dig);
1367 ###### number includes
1370 To parse a full number we need to consider the optional base, the
1371 mantissa, and the optional exponent. We will treat these one at a
1374 The base is indicated by a letter after a leading zero, which must be
1375 followed by a base letter or a period. The base also determines the
1376 character which will mark an exponent.
1384 if (tok.txt[0] == '0' && tok.len > 1) {
1386 switch(tok.txt[1]) {
1417 // another digit is not permitted
1421 // must be decimal marker or trailing
1422 // letter, which are OK;
1429 After the base is the mantissa, which may contain a decimal mark, so
1430 we need to record the number of places. We won't impose the number of
1431 places until we have the exponent as well.
1438 ###### parse mantissa
1440 d = parse_digits(mant, tok, base, &places);
1446 mpq_set_z(num, mant);
1449 After the mantissa number may come an exponent which may be positive
1450 or negative. We assume at this point that we have seen the exponent
1458 ###### parse exponent
1460 if (tok.txt[0] == '+') {
1463 } else if (tok.txt[0] == '-') {
1469 d = parse_digits(exp, tok, 10, NULL);
1474 if (!mpz_fits_slong_p(exp)) {
1479 lexp = mpz_get_si(exp) * esign;
1485 Now that we have the mantissa and the exponent we can multiply them
1486 together, also allowing for the number of digits after the decimal
1489 For base 10, we simply subtract the decimal places from the exponent.
1490 For the other bases, as the exponent is alway based on 2, even for
1491 octal and hex, we need a bit more detail.
1492 We then recover the sign from the exponent, as division is quite
1493 different from multiplication.
1495 ###### calc exponent
1514 Imposing the exponent on the number is also very different for base 10
1515 than for the others. For the binary shift `gmp` provides a simple
1516 function. For base 10 we use something like Russian Peasant
1519 ###### calc exponent
1523 mpq_set_ui(tens, 10, 1);
1527 mpq_mul(num, num, tens);
1529 mpq_div(num, num, tens);
1534 mpq_mul(tens, tens, tens);
1539 mpq_mul_2exp(num, num, lexp);
1541 mpq_div_2exp(num, num, lexp);
1544 Now we are ready to parse a number: the base, mantissa, and exponent.
1545 If all goes well we check for the possible trailing letters and
1546 return. Return value is 1 for success and 0 for failure.
1549 ###### number functions
1550 int number_parse(mpq_t num, char tail[3], struct text tok)
1557 if (tok.len > 1 && (tok.txt[0] == expc ||
1558 tok.txt[0] == toupper(expc))) {
1565 for (i = 0; i < 2; i++) {
1568 if (!isalpha(tok.txt[i]))
1570 tail[i] = tok.txt[i];
1580 Number parsing goes in `libnumber.c`
1582 ###### File: libnumber.c
1590 ###### File: number.h
1591 int number_parse(mpq_t num, char tail[3], struct text tok);
1593 ###### File: scanner.mk
1595 libnumber.o : libnumber.c
1596 $(CC) $(CFLAGS) -c libnumber.c
1598 ## Processing strings
1600 Both `TK_string` and `TK_multi_string` require post-processing which
1601 can be one of two types: literal or with escapes processed.
1602 Even literal processing is non-trivial as the file may contain indents
1603 which need to be stripped.
1605 Errors can only occur when processing escapes. Any unrecognised
1606 character following the escape character will cause an error.
1608 Processing escapes and striping indents can only make the string
1609 shorter, not longer, so we allocate a buffer which is the same size as
1610 the string and process into that.
1612 To request escape processing, we pass the character we want to use for
1613 quoting, usually '`\`'. To avoid escape processing we pass a zero.
1616 int string_parse(struct token *tok, char escape,
1617 struct text *str, char tail[3])
1620 struct text t = tok->txt;
1624 if (tok->num == TK_string) {
1629 str->txt = malloc(t.len);
1642 The tail of the string can be 0, 1, or 2 letters
1645 if (i >= 0 && isalpha(t.txt[i-1]))
1647 if (i >= 0 && isalpha(t.txt[i-1]))
1649 strncpy(tail, t.txt+i, t.len-i);
1658 Stripping the quote of a single-line string is trivial.
1659 The only part that is at all interesting is that quote character must
1663 if (t.txt[t.len-1] != quote)
1673 For a multi-line string we have a little more work to do. We need to
1674 remove 3 quotes, not 1, and need to count the indent of the close
1675 quote as it will need to be stripped from all lines.
1679 t.txt[1] != quote || t.txt[2] != quote ||
1680 !is_newline(t.txt[3]))
1685 if (i <= 0 || t.txt[i-1] != quote)
1688 if (i <= 0 || t.txt[i-1] != quote)
1691 if (i <= 0 || t.txt[i-1] != quote)
1695 while (i > 0 && !is_newline(t.txt[i-1]))
1699 if (t.txt[i] == ' ')
1701 if (t.txt[i] == '\t')
1702 indent = indent_tab(indent);
1711 Now we just take one byte at a time. trans-ASCII unicode won't look
1712 like anything we are interested in so it will just be copied byte by
1717 for (i = 0; i < t.len; i++) {
1731 } else if (i+1 >= t.len) {
1732 // escape and end of string
1740 str->len = cp - str->txt;
1748 Every time we find a start of line, we strip spaces and tabs until the
1749 required indent is found.
1752 while (i < t.len && skipped < indent) {
1757 skipped = indent_tab(c);
1766 *cp++ = '\n'; break;
1768 *cp++ = '\r'; break;
1770 *cp++ = '\t'; break;
1772 *cp++ = '\b'; break;
1774 *cp++ = quote; break;
1776 *cp++ = '\f'; break;
1778 *cp++ = '\v'; break;
1780 *cp++ = '\a'; break;
1785 // 3 digit octal number
1788 if (t.txt[i+1] < '0' || t.txt[i+1] > '7' ||
1789 t.txt[i+2] < '0' || t.txt[i+1] > '7')
1791 n = (t.txt[i ]-'0') * 64 +
1792 (t.txt[i+1]-'0') * 8 +
1793 (t.txt[i+2]-'0') * 1;
1799 n = take_hex(2, t.txt+i+1, t.len-i-1);
1807 // 4 or 8 hex digits for unicode
1808 n = take_hex(c == 'u'?4:8, t.txt+i+1, t.len-i-1);
1811 memset(&pstate, 0, sizeof(pstate));
1812 n = wcrtomb(cp, n, &pstate);
1816 i += c == 'u' ? 4 : 8;
1821 else if (is_newline(c))
1831 For `\x` `\u` and `\U` we need to collect a specific number of
1834 ###### string functions
1836 static long take_hex(int digits, char *cp, int l)
1848 else if (isupper(c))
1859 #### File: libstring.c
1861 String parsing goes in `libstring.c`
1870 #include "scanner.h"
1874 ###### File: string.h
1875 int string_parse(struct token *tok, char escape,
1876 struct text *str, char tail[3]);
1878 ###### File: scanner.mk
1880 libstring.o : libstring.c
1881 $(CC) $(CFLAGS) -c libstring.c
1886 As "untested code is buggy code" we need a program to easily test
1887 the scanner library. This will simply parse a given file and report
1888 the tokens one per line.
1890 ###### File: scanner.c
1896 #include <sys/mman.h>
1902 #include "scanner.h"
1907 static void pr_err(char *msg)
1910 fprintf(stderr, "%s\n", msg);
1913 int main(int argc, char *argv[])
1918 struct token_state *state;
1919 const char *known[] = {
1928 struct token_config conf = {
1931 .words_marks = known,
1932 .number_chars = "., _+-",
1933 .known_count = sizeof(known)/sizeof(known[0]),
1934 .ignored = (0 << TK_line_comment)
1935 |(0 << TK_block_comment),
1937 struct section *table, *s, *prev;
1938 setlocale(LC_ALL,"");
1940 fprintf(stderr, "Usage: scanner file\n");
1943 fd = open(argv[1], O_RDONLY);
1945 fprintf(stderr, "scanner: cannot open %s: %s\n",
1946 argv[1], strerror(errno));
1949 len = lseek(fd, 0, 2);
1950 file = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0);
1951 table = code_extract(file, file+len, pr_err);
1954 (code_free(s->code), prev = s, s = s->next, free(prev))) {
1955 printf("Tokenizing: %.*s\n", s->section.len,
1957 state = token_open(s->code, &conf);
1959 struct token tk = token_next(state);
1960 printf("%d:%d ", tk.line, tk.col);
1961 token_trace(stdout, tk, 20);
1962 if (tk.num == TK_number) {
1965 if (number_parse(num, tail,tk.txt)) {
1966 printf(" %s ", tail);
1967 mpq_out_str(stdout, 10, num);
1970 printf(" BAD NUMBER");
1972 if (tk.num == TK_string ||
1973 tk.num == TK_multi_string) {
1977 if (tk.txt.txt[0] == '`')
1979 if (string_parse(&tk, esc,
1981 printf(" %s ", tail);
1982 text_dump(stdout, str, 20);
1985 printf(" BAD STRING");
1988 if (tk.num == TK_error)
1990 if (tk.num == TK_eof)
1996 ###### File: scanner.mk
1997 scanner.c : scanner.mdc
2000 scanner : scanner.o scanner.h libscanner.o libmdcode.o mdcode.h
2001 $(CC) $(CFLAGS) -o scanner scanner.o libscanner.o \
2002 libmdcode.o libnumber.o libstring.o -licuuc -lgmp
2003 scanner.o : scanner.c
2004 $(CC) $(CFLAGS) -c scanner.c