1 # LR(1) Parser Generator #
3 This parser generator takes a grammar description combined with code
4 fragments, analyses it, and can report details about the analysis and
5 write out C code files which can be compiled to make a parser.
7 There are several distinct sections.
9 1. `grammar_read` will read a grammar from a literate-code file and
10 store the productions, symbols, and code fragments.
11 2. `grammar_analyse` will take that grammar and build LR parsing
12 states and look-ahead tables.
13 3. `grammar_report` will present the details of the analysis
14 in a readable format and will report any conflicts.
15 4. `parser_generate` will write out C code files with various
16 tables and with the code fragments arranged in useful places.
17 5. `parser_run` is a library function intended to be linked together
18 with the generated parser tables to complete the implementation of
20 6. Finally `calc` is a test grammar for a simple calculator. The
21 `parsergen` program built from the C code in this file can extract
22 that grammar directly from this file and process it.
25 ###### File: parsergen.c
30 ## forward declarations
41 ###### File: libparser.c
48 ###### File: parsergen.mk
51 parsergen.c parsergen.mk libparser.c parser.h : parsergen.mdc
54 ## Reading the grammar
56 The grammar must be stored in a markdown literate code file as parsed
57 by "[mdcode][]". It must have three top level (i.e. unreferenced)
58 sections: `header`, `code`, and `grammar`. The first two will be
59 literally copied into the generated `.c` and `.h`. files. The last
60 contains the grammar. This is tokenised with "[scanner][]".
62 If the `--tag` option is given, then any top level header that doesn't
63 start with the tag is ignored, and the tag is striped from the rest. So
65 means that the three needed sections must be `Foo: header`, `Foo: code`,
69 [scanner]: scanner.html
75 ###### parser includes
79 The grammar contains production sets, precedence/associativity
80 declarations, and data type declarations. These are all parsed with
81 _ad hoc_ parsing as we don't have a parser generator yet.
83 The precedence and associativity can be set for each production, but
84 can be inherited from symbols. The data type (either structure or a
85 reference to a structure) is potentially defined for each non-terminal
86 and describes what C structure is used to store information about each
90 enum assoc {Left, Right, Non};
91 char *assoc_names[] = {"Left","Right","Non"};
94 struct text struct_name;
97 unsigned short precedence;
101 unsigned short precedence;
109 The strings reported by `mdcode` and `scanner` are `struct text` which have
110 length rather than being null terminated. To help with printing and
111 comparing we define `text_is` and `prtxt`, which should possibly go in
112 `mdcode`. `scanner` does provide `text_dump` which is useful for strings
113 which might contain control characters.
115 `strip_tag` is a bit like `strncmp`, but adds a test for a colon,
116 because that is what we need to detect tags.
119 static int text_is(struct text t, char *s)
121 return (strlen(s) == t.len &&
122 strncmp(s, t.txt, t.len) == 0);
124 static void prtxt(struct text t)
126 printf("%.*s", t.len, t.txt);
129 static int strip_tag(struct text *t, char *tag)
131 int skip = strlen(tag) + 1;
132 if (skip >= t->len ||
133 strncmp(t->txt, tag, skip-1) != 0 ||
134 t->txt[skip-1] != ':')
136 while (skip < t->len && t->txt[skip] == ' ')
145 Productions are comprised primarily of symbols - terminal and
146 non-terminal. We do not make a syntactic distinction between the two,
147 though non-terminals must be identifiers. Non-terminal symbols are
148 those which appear in the head of a production, terminal symbols are
149 those which don't. There are also "virtual" symbols used for precedence
150 marking discussed later, and sometimes we won't know what type a symbol
153 ###### forward declarations
154 enum symtype { Unknown, Virtual, Terminal, Nonterminal };
155 char *symtypes = "UVTN";
159 Symbols can be either `TK_ident` or `TK_mark`. They are saved in a
160 table of known symbols and the resulting parser will report them as
161 `TK_reserved + N`. A small set of identifiers are reserved for the
162 different token types that `scanner` can report.
165 static char *reserved_words[] = {
166 [TK_error] = "ERROR",
167 [TK_number] = "NUMBER",
168 [TK_ident] = "IDENTIFIER",
170 [TK_string] = "STRING",
171 [TK_multi_string] = "MULTI_STRING",
174 [TK_newline] = "NEWLINE",
180 Note that `TK_eof` and the two `TK_*_comment` tokens cannot be
181 recognised. The former is automatically expected at the end of the text
182 being parsed. The latter are always ignored by the parser.
184 All of these symbols are stored in a simple symbol table. We use the
185 `struct text` from `mdcode` to store the name and link them together
186 into a sorted list using an insertion sort.
188 We don't have separate `find` and `insert` functions as any symbol we
189 find needs to be remembered. We simply expect `find` to always return a
190 symbol, but its type might be `Unknown`.
199 ###### grammar fields
204 static struct symbol *sym_find(struct grammar *g, struct text s)
206 struct symbol **l = &g->syms;
211 (cmp = text_cmp((*l)->name, s)) < 0)
215 n = calloc(1, sizeof(*n));
224 static void symbols_init(struct grammar *g)
226 int entries = sizeof(reserved_words)/sizeof(reserved_words[0]);
228 for (i = 0; i < entries; i++) {
231 t.txt = reserved_words[i];
234 t.len = strlen(t.txt);
241 ### Data types and precedence.
243 Data type specification and precedence specification are both
244 introduced by a dollar sign at the start of the line. If the next
245 word is `LEFT`, `RIGHT` or `NON`, then the line specifies a
246 precedence, otherwise it specifies a data type.
248 The data type name is simply stored and applied to the head of all
249 subsequent productions. It must be the name of a structure optionally
250 preceded by an asterisk which means a reference or "pointer". So
251 `$expression` maps to `struct expression` and `$*statement` maps to
252 `struct statement *`.
254 Any productions given before the first data type declaration will have
255 no data type associated with them and can carry no information. In
256 order to allow other non-terminals to have no type, the data type
257 `$void` can be given. This does *not* mean that `struct void` will be
258 used, but rather than no type will be associated with future
261 The precedence line must contain a list of symbols - typically
262 terminal symbols, but not necessarily. It can only contain symbols
263 that have not been seen yet, so precedence declaration must precede
264 the productions that they affect.
266 A precedence line may also contain a "virtual" symbol which is an
267 identifier preceded by `$$`. Virtual terminals carry precedence
268 information but are not included in the grammar. A production can
269 declare that it inherits the precedence of a given virtual symbol.
271 This use for `$$` precludes it from being used as a symbol in the
272 described language. Two other symbols: `${` and `}$` are also
275 Each new precedence line introduces a new precedence level and
276 declares how it associates. This level is stored in each symbol
277 listed and may be inherited by any production which uses the symbol. A
278 production inherits from the last symbol which has a precedence.
280 ###### grammar fields
281 struct text current_type;
286 enum symbols { TK_virtual = TK_reserved, TK_open, TK_close };
287 static const char *known[] = { "$$", "${", "}$" };
290 static char *dollar_line(struct token_state *ts, struct grammar *g, int isref)
292 struct token t = token_next(ts);
297 if (t.num != TK_ident) {
298 err = "type or assoc expected after '$'";
301 if (text_is(t.txt, "LEFT"))
303 else if (text_is(t.txt, "RIGHT"))
305 else if (text_is(t.txt, "NON"))
308 g->current_type = t.txt;
309 g->type_isref = isref;
310 if (text_is(t.txt, "void"))
311 g->current_type.txt = NULL;
313 if (t.num != TK_newline) {
314 err = "Extra tokens after type name";
321 err = "$* cannot be followed by a precedence";
325 // This is a precedence line, need some symbols.
329 while (t.num != TK_newline) {
330 enum symtype type = Terminal;
332 if (t.num == TK_virtual) {
335 if (t.num != TK_ident) {
336 err = "$$ must be followed by a word";
339 } else if (t.num != TK_ident &&
341 err = "Illegal token in precedence line";
344 s = sym_find(g, t.txt);
345 if (s->type != Unknown) {
346 err = "Symbols in precedence line must not already be known.";
350 s->precedence = g->prec_levels;
355 err = "No symbols given on precedence line";
359 while (t.num != TK_newline && t.num != TK_eof)
366 A production either starts with an identifier which is the head
367 non-terminal, or a vertical bar (`|`) in which case this production
368 uses the same head as the previous one. The identifier must be
369 followed by a `->` mark. All productions for a given non-terminal must
370 be grouped together with the `nonterminal ->` given only once.
372 After this a (possibly empty) sequence of identifiers and marks form
373 the body of the production. A virtual symbol may be given after the
374 body by preceding it with `$$`. If a virtual symbol is given, the
375 precedence of the production is that for the virtual symbol. If none
376 is given, the precedence is inherited from the last symbol in the
377 production which has a precedence specified.
379 After the optional precedence may come the `${` mark. This indicates
380 the start of a code fragment. If present, this must be on the same
381 line as the start of the production.
383 All of the text from the `${` through to the matching `}$` is
384 collected and forms the code-fragment for the production. It must all
385 be in one `code_node` of the literate code. The `}$` must be
386 at the end of a line.
388 Text in the code fragment will undergo substitutions where `$N` or
389 `$<N`,for some numeric `N`, will be replaced with a variable holding
390 the parse information for the particular symbol in the production.
391 `$0` is the head of the production, `$1` is the first symbol of the
392 body, etc. The type of `$N` for a terminal symbol is `struct token`.
393 For a non-terminal, it is whatever has been declared for that symbol.
394 The `<` may be included for symbols declared as storing a reference
395 (not a structure) and means that the reference is being moved out, so
396 it will not automatically be freed.
398 While building productions we will need to add to an array which needs to
402 static void array_add(void *varray, int *cnt, void *new)
404 void ***array = varray;
407 current = ((*cnt-1) | (step-1))+1;
408 if (*cnt == current) {
411 *array = realloc(*array, current * sizeof(void*));
413 (*array)[*cnt] = new;
417 Collecting the code fragment simply involves reading tokens until we
418 hit the end token `}$`, and noting the character position of the start and
422 static struct text collect_code(struct token_state *state,
427 code.txt = start.txt.txt + start.txt.len;
429 t = token_next(state);
430 while (t.node == start.node &&
431 t.num != TK_close && t.num != TK_error &&
433 if (t.num == TK_close && t.node == start.node)
434 code.len = t.txt.txt - code.txt;
440 Now we have all the bit we need to parse a full production.
445 ###### grammar fields
446 struct production **productions;
447 int production_count;
449 ###### production fields
451 struct symbol **body;
456 int first_production;
459 static char *parse_production(struct grammar *g,
461 struct token_state *state)
463 /* Head has already been parsed. */
466 struct production p, *pp;
468 memset(&p, 0, sizeof(p));
470 tk = token_next(state);
471 while (tk.num == TK_ident || tk.num == TK_mark) {
472 struct symbol *bs = sym_find(g, tk.txt);
473 if (bs->type == Unknown)
475 if (bs->type == Virtual) {
476 err = "Virtual symbol not permitted in production";
479 if (bs->precedence) {
480 p.precedence = bs->precedence;
483 array_add(&p.body, &p.body_size, bs);
484 tk = token_next(state);
486 if (tk.num == TK_virtual) {
488 tk = token_next(state);
489 if (tk.num != TK_ident) {
490 err = "word required after $$";
493 vs = sym_find(g, tk.txt);
494 if (vs->type != Virtual) {
495 err = "symbol after $$ must be virtual";
498 p.precedence = vs->precedence;
500 tk = token_next(state);
502 if (tk.num == TK_open) {
503 p.code = collect_code(state, tk);
504 if (p.code.txt == NULL) {
505 err = "code fragment not closed properly";
508 tk = token_next(state);
510 if (tk.num != TK_newline && tk.num != TK_eof) {
511 err = "stray tokens at end of line";
514 pp = malloc(sizeof(*pp));
516 array_add(&g->productions, &g->production_count, pp);
519 while (tk.num != TK_newline && tk.num != TK_eof)
520 tk = token_next(state);
524 With the ability to parse production and dollar-lines, we have nearly all
525 that we need to parse a grammar from a `code_node`.
527 The head of the first production will effectively be the `start` symbol of
528 the grammar. However it won't _actually_ be so. Processing the grammar is
529 greatly simplified if the real start symbol only has a single production,
530 and expects `$eof` as the final terminal. So when we find the first
531 explicit production we insert an extra production as production zero which
534 ###### Example: production 0
537 where `START` is the first non-terminal given.
539 ###### create production zero
540 struct production *p = calloc(1,sizeof(*p));
541 struct text start = {"$start",6};
542 struct text eof = {"$eof",4};
543 struct text code = {"$0 = $<1;", 9};
544 p->head = sym_find(g, start);
545 p->head->type = Nonterminal;
546 p->head->struct_name = g->current_type;
547 p->head->isref = g->type_isref;
548 if (g->current_type.txt)
550 array_add(&p->body, &p->body_size, head);
551 array_add(&p->body, &p->body_size, sym_find(g, eof));
552 p->head->first_production = g->production_count;
553 array_add(&g->productions, &g->production_count, p);
555 Now we are ready to read in the grammar.
558 static struct grammar *grammar_read(struct code_node *code)
560 struct token_config conf = {
563 .words_marks = known,
564 .known_count = sizeof(known)/sizeof(known[0]),
566 .ignored = (1 << TK_line_comment)
567 | (1 << TK_block_comment)
570 | (1 << TK_multi_string)
575 struct token_state *state = token_open(code, &conf);
577 struct symbol *head = NULL;
581 g = calloc(1, sizeof(*g));
584 for (tk = token_next(state); tk.num != TK_eof;
585 tk = token_next(state)) {
586 if (tk.num == TK_newline)
588 if (tk.num == TK_ident) {
590 head = sym_find(g, tk.txt);
591 if (head->type == Nonterminal)
592 err = "This non-terminal has already be used.";
593 else if (head->type == Virtual)
594 err = "Virtual symbol not permitted in head of production";
596 head->type = Nonterminal;
597 head->struct_name = g->current_type;
598 head->isref = g->type_isref;
599 if (g->production_count == 0) {
600 ## create production zero
602 head->first_production = g->production_count;
603 tk = token_next(state);
604 if (tk.num == TK_mark &&
605 text_is(tk.txt, "->"))
606 err = parse_production(g, head, state);
608 err = "'->' missing in production";
610 } else if (tk.num == TK_mark
611 && text_is(tk.txt, "|")) {
612 // another production for same non-term
614 err = parse_production(g, head, state);
616 err = "First production must have a head";
617 } else if (tk.num == TK_mark
618 && text_is(tk.txt, "$")) {
619 err = dollar_line(state, g, 0);
620 } else if (tk.num == TK_mark
621 && text_is(tk.txt, "$*")) {
622 err = dollar_line(state, g, 1);
624 err = "Unrecognised token at start of line.";
632 fprintf(stderr, "Error at line %d: %s\n",
639 ## Analysing the grammar
641 The central task in analysing the grammar is to determine a set of
642 states to drive the parsing process. This will require finding
643 various sets of symbols and of "items". Some of these sets will need
644 to attach information to each element in the set, so they are more
647 Each "item" may have a 'look-ahead' or `LA` set associated with
648 it. Many of these will be the same as each other. To avoid memory
649 wastage, and to simplify some comparisons of sets, these sets will be
650 stored in a list of unique sets, each assigned a number.
652 Once we have the data structures in hand to manage these sets and
653 lists, we can start setting the 'nullable' flag, build the 'FIRST'
654 sets, and then create the item sets which define the various states.
658 Though we don't only store symbols in these sets, they are the main
659 things we store, so they are called symbol sets or "symsets".
661 A symset has a size, an array of shorts, and an optional array of data
662 values, which are also shorts. If the array of data is not present,
663 we store an impossible pointer, as `NULL` is used to indicate that no
664 memory has been allocated yet;
669 unsigned short *syms, *data;
671 #define NO_DATA ((unsigned short *)1)
672 const struct symset INIT_SYMSET = { 0, NULL, NO_DATA };
673 const struct symset INIT_DATASET = { 0, NULL, NULL };
675 The arrays of shorts are allocated in blocks of 8 and are kept sorted
676 by using an insertion sort. We don't explicitly record the amount of
677 allocated space as it can be derived directly from the current `cnt` using
678 `((cnt - 1) | 7) + 1`.
681 static void symset_add(struct symset *s, unsigned short key, unsigned short val)
684 int current = ((s->cnt-1) | 7) + 1;
685 if (current == s->cnt) {
687 s->syms = realloc(s->syms, sizeof(*s->syms) * current);
688 if (s->data != NO_DATA)
689 s->data = realloc(s->data, sizeof(*s->data) * current);
692 while (i > 0 && s->syms[i-1] > key) {
693 s->syms[i] = s->syms[i-1];
694 if (s->data != NO_DATA)
695 s->data[i] = s->data[i-1];
699 if (s->data != NO_DATA)
704 Finding a symbol (or item) in a `symset` uses a simple binary search.
705 We return the index where the value was found (so data can be accessed),
706 or `-1` to indicate failure.
708 static int symset_find(struct symset *ss, unsigned short key)
715 while (lo + 1 < hi) {
716 int mid = (lo + hi) / 2;
717 if (ss->syms[mid] <= key)
722 if (ss->syms[lo] == key)
727 We will often want to form the union of two symsets. When we do, we
728 will often want to know if anything changed (as they might mean there
729 is more work to do). So `symset_union` reports whether anything was
730 added to the first set. We use a slow quadratic approach as these
731 sets don't really get very big. If profiles shows this to be a problem is
732 can be optimised later.
734 static int symset_union(struct symset *a, struct symset *b)
738 for (i = 0; i < b->cnt; i++)
739 if (symset_find(a, b->syms[i]) < 0) {
740 unsigned short data = 0;
741 if (b->data != NO_DATA)
743 symset_add(a, b->syms[i], data);
749 And of course we must be able to free a symset.
751 static void symset_free(struct symset ss)
754 if (ss.data != NO_DATA)
760 Some symsets are simply stored somewhere appropriate in the `grammar`
761 data structure, others needs a bit of indirection. This applies
762 particularly to "LA" sets. These will be paired with "items" in an
763 "itemset". As itemsets will be stored in a symset, the "LA" set needs to be
764 stored in the `data` field, so we need a mapping from a small (short)
765 number to an LA `symset`.
767 As mentioned earlier this involves creating a list of unique symsets.
769 For now, we just use a linear list sorted by insertion. A skiplist
770 would be more efficient and may be added later.
777 struct setlist *next;
780 ###### grammar fields
781 struct setlist *sets;
786 static int ss_cmp(struct symset a, struct symset b)
789 int diff = a.cnt - b.cnt;
792 for (i = 0; i < a.cnt; i++) {
793 diff = (int)a.syms[i] - (int)b.syms[i];
800 static int save_set(struct grammar *g, struct symset ss)
802 struct setlist **sl = &g->sets;
806 while (*sl && (cmp = ss_cmp((*sl)->ss, ss)) < 0)
813 s = malloc(sizeof(*s));
822 Finding a set by number is currently performed by a simple linear search.
823 If this turns out to hurt performance, we can store the sets in a dynamic
824 array like the productions.
826 static struct symset set_find(struct grammar *g, int num)
828 struct setlist *sl = g->sets;
829 while (sl && sl->num != num)
835 ### Setting `nullable`
837 We set `nullable` on the head symbol for any production for which all
838 the body symbols (if any) are nullable. As this is a recursive
839 definition, any change in the `nullable` setting means that we need to
840 re-evaluate where it needs to be set.
842 We simply loop around performing the same calculations until no more
849 static void set_nullable(struct grammar *g)
852 while (check_again) {
855 for (p = 0; p < g->production_count; p++) {
856 struct production *pr = g->productions[p];
859 if (pr->head->nullable)
861 for (s = 0; s < pr->body_size; s++)
862 if (! pr->body[s]->nullable)
864 if (s == pr->body_size) {
865 pr->head->nullable = 1;
872 ### Setting `can_eol`
874 In order to be able to ignore newline tokens when not relevant, but
875 still include them in the parse when needed, we will need to know
876 which states can start a "line-like" section of code. We ignore
877 newlines when there is an indent since the most recent start of a
880 To know what is line-like, we first need to know which symbols can end
881 a line-like section, which is precisely those which can end with a
882 newline token. These symbols don't necessarily alway end with a
883 newline, but they can. Hence they are not described as "lines" but
886 Clearly the `TK_newline` token can end with a newline. Any symbol
887 which is the head of a production that contains a line-ending symbol
888 followed only by nullable symbols is also a line-ending symbol. We
889 use a new field `can_eol` to record this attribute of symbols, and
890 compute it in a repetitive manner similar to `set_nullable`.
896 static void set_can_eol(struct grammar *g)
899 g->symtab[TK_newline]->can_eol = 1;
900 while (check_again) {
903 for (p = 0; p < g->production_count; p++) {
904 struct production *pr = g->productions[p];
907 if (pr->head->can_eol)
910 for (s = pr->body_size - 1; s >= 0; s--) {
911 if (pr->body[s]->can_eol) {
912 pr->head->can_eol = 1;
916 if (!pr->body[s]->nullable)
923 ### Building the `first` sets
925 When calculating what can follow a particular non-terminal, we will need to
926 know what the "first" terminal in any subsequent non-terminal might be. So
927 we calculate the `first` set for every non-terminal and store them in an
928 array. We don't bother recording the "first" set for terminals as they are
931 As the "first" for one symbol might depend on the "first" of another,
932 we repeat the calculations until no changes happen, just like with
933 `set_nullable`. This makes use of the fact that `symset_union`
934 reports if any change happens.
936 The core of this which finds the "first" of part of a production body
937 will be reused for computing the "follow" sets, so we split it out
938 into a separate function.
940 ###### grammar fields
941 struct symset *first;
945 static int add_first(struct production *pr, int start,
946 struct symset *target, struct grammar *g,
951 for (s = start; s < pr->body_size; s++) {
952 struct symbol *bs = pr->body[s];
953 if (bs->type == Terminal) {
954 if (symset_find(target, bs->num) < 0) {
955 symset_add(target, bs->num, 0);
959 } else if (symset_union(target, &g->first[bs->num]))
965 *to_end = (s == pr->body_size);
969 static void build_first(struct grammar *g)
973 g->first = calloc(g->num_syms, sizeof(g->first[0]));
974 for (s = 0; s < g->num_syms; s++)
975 g->first[s] = INIT_SYMSET;
977 while (check_again) {
980 for (p = 0; p < g->production_count; p++) {
981 struct production *pr = g->productions[p];
982 struct symset *head = &g->first[pr->head->num];
984 if (add_first(pr, 0, head, g, NULL))
990 ### Building the `follow` sets.
992 There are two different situations when we will want to generate "follow"
993 sets. If we are doing an SLR analysis, we want to generate a single
994 "follow" set for each non-terminal in the grammar. That is what is
995 happening here. If we are doing an LALR or LR analysis we will want
996 to generate a separate "LA" set for each item. We do that later
999 There are two parts to generating a "follow" set. Firstly we look at
1000 every place that any non-terminal appears in the body of any
1001 production, and we find the set of possible "first" symbols after
1002 there. This is done using `add_first` above and only needs to be done
1003 once as the "first" sets are now stable and will not change.
1007 for (p = 0; p < g->production_count; p++) {
1008 struct production *pr = g->productions[p];
1011 for (b = 0; b < pr->body_size - 1; b++) {
1012 struct symbol *bs = pr->body[b];
1013 if (bs->type == Terminal)
1015 add_first(pr, b+1, &g->follow[bs->num], g, NULL);
1019 The second part is to add the "follow" set of the head of a production
1020 to the "follow" sets of the final symbol in the production, and any
1021 other symbol which is followed only by `nullable` symbols. As this
1022 depends on "follow" itself we need to repeatedly perform the process
1023 until no further changes happen.
1027 for (again = 0, p = 0;
1028 p < g->production_count;
1029 p < g->production_count-1
1030 ? p++ : again ? (again = 0, p = 0)
1032 struct production *pr = g->productions[p];
1035 for (b = pr->body_size - 1; b >= 0; b--) {
1036 struct symbol *bs = pr->body[b];
1037 if (bs->type == Terminal)
1039 if (symset_union(&g->follow[bs->num],
1040 &g->follow[pr->head->num]))
1047 We now just need to create and initialise the `follow` list to get a
1050 ###### grammar fields
1051 struct symset *follow;
1054 static void build_follow(struct grammar *g)
1057 g->follow = calloc(g->num_syms, sizeof(g->follow[0]));
1058 for (s = 0; s < g->num_syms; s++)
1059 g->follow[s] = INIT_SYMSET;
1063 ### Building itemsets and states
1065 There are three different levels of detail that can be chosen for
1066 building the itemsets and states for the LR grammar. They are:
1068 1. LR(0) or SLR(1), where no look-ahead is considered.
1069 2. LALR(1) where we build look-ahead sets with each item and merge
1070 the LA sets when we find two paths to the same "kernel" set of items.
1071 3. LR(1) where different look-ahead for any item in the set means
1072 a different state must be created.
1074 ###### forward declarations
1075 enum grammar_type { LR0, LR05, SLR, LALR, LR1 };
1077 We need to be able to look through existing states to see if a newly
1078 generated state already exists. For now we use a simple sorted linked
1081 An item is a pair of numbers: the production number and the position of
1082 "DOT", which is an index into the body of the production.
1083 As the numbers are not enormous we can combine them into a single "short"
1084 and store them in a `symset` - 4 bits for "DOT" and 12 bits for the
1085 production number (so 4000 productions with maximum of 15 symbols in the
1088 Comparing the itemsets will be a little different to comparing symsets
1089 as we want to do the lookup after generating the "kernel" of an
1090 itemset, so we need to ignore the offset=zero items which are added during
1093 To facilitate this, we modify the "DOT" number so that "0" sorts to
1094 the end of the list in the symset, and then only compare items before
1098 static inline unsigned short item_num(int production, int index)
1100 return production | ((31-index) << 11);
1102 static inline int item_prod(unsigned short item)
1104 return item & 0x7ff;
1106 static inline int item_index(unsigned short item)
1108 return (31-(item >> 11)) & 0x1f;
1111 For LR(1) analysis we need to compare not just the itemset in a state
1112 but also the LA sets. As we assign each unique LA set a number, we
1113 can just compare the symset and the data values together.
1116 static int itemset_cmp(struct symset a, struct symset b,
1117 enum grammar_type type)
1123 i < a.cnt && i < b.cnt &&
1124 item_index(a.syms[i]) > 0 &&
1125 item_index(b.syms[i]) > 0;
1127 int diff = a.syms[i] - b.syms[i];
1131 diff = a.data[i] - b.data[i];
1136 if (i == a.cnt || item_index(a.syms[i]) == 0)
1140 if (i == b.cnt || item_index(b.syms[i]) == 0)
1146 if (type < LR1 || av == -1)
1149 a.data[i] - b.data[i];
1152 And now we can build the list of itemsets. The lookup routine returns
1153 both a success flag and a pointer to where in the list an insert
1154 should happen, so we don't need to search a second time.
1158 struct itemset *next;
1160 struct symset items;
1161 struct symset go_to;
1166 ###### grammar fields
1167 struct itemset *items;
1171 static int itemset_find(struct grammar *g, struct itemset ***where,
1172 struct symset kernel, enum grammar_type type)
1174 struct itemset **ip;
1176 for (ip = &g->items; *ip ; ip = & (*ip)->next) {
1177 struct itemset *i = *ip;
1179 diff = itemset_cmp(i->items, kernel, type);
1192 Adding an itemset may require merging the LA sets if LALR analysis is
1193 happening. If any new LA set adds any symbols that weren't in the old LA set, we
1194 clear the `completed` flag so that the dependants of this itemset will be
1195 recalculated and their LA sets updated.
1197 `add_itemset` must consume the symsets it is passed, either by adding
1198 them to a data structure, of freeing them.
1200 static int add_itemset(struct grammar *g, struct symset ss,
1201 enum grammar_type type, int starts_line)
1203 struct itemset **where, *is;
1205 int found = itemset_find(g, &where, ss, type);
1207 is = calloc(1, sizeof(*is));
1208 is->state = g->states;
1212 is->go_to = INIT_DATASET;
1213 is->starts_line = starts_line;
1222 for (i = 0; i < ss.cnt; i++) {
1223 struct symset temp = INIT_SYMSET, s;
1224 if (ss.data[i] == is->items.data[i])
1226 s = set_find(g, is->items.data[i]);
1227 symset_union(&temp, &s);
1228 s = set_find(g, ss.data[i]);
1229 if (symset_union(&temp, &s)) {
1230 is->items.data[i] = save_set(g, temp);
1241 To build all the itemsets, we first insert the initial itemset made
1242 from production zero, complete each itemset, and then generate new
1243 itemsets from old until no new ones can be made.
1245 Completing an itemset means finding all the items where "DOT" is followed by
1246 a nonterminal and adding "DOT=0" items for every production from that
1247 non-terminal - providing each item hasn't already been added.
1249 If LA sets are needed, the LA set for each new item is found using
1250 `add_first` which was developed earlier for `FIRST` and `FOLLOW`. This may
1251 be supplemented by the LA set for the item which produce the new item.
1253 We also collect a set of all symbols which follow "DOT" (in `done`) as this
1254 is used in the next stage.
1256 NOTE: precedence handling should happen here - I haven't written this yet
1259 ###### complete itemset
1260 for (i = 0; i < is->items.cnt; i++) {
1261 int p = item_prod(is->items.syms[i]);
1262 int bs = item_index(is->items.syms[i]);
1263 struct production *pr = g->productions[p];
1266 struct symset LA = INIT_SYMSET;
1267 unsigned short sn = 0;
1269 if (bs == pr->body_size)
1272 if (symset_find(&done, s->num) < 0)
1273 symset_add(&done, s->num, 0);
1274 if (s->type != Nonterminal)
1280 add_first(pr, bs+1, &LA, g, &to_end);
1282 struct symset ss = set_find(g, is->items.data[i]);
1283 symset_union(&LA, &ss);
1285 sn = save_set(g, LA);
1286 LA = set_find(g, sn);
1289 /* Add productions for this symbol */
1290 for (p2 = s->first_production;
1291 p2 < g->production_count &&
1292 g->productions[p2]->head == s;
1294 int itm = item_num(p2, 0);
1295 int pos = symset_find(&is->items, itm);
1297 symset_add(&is->items, itm, sn);
1298 /* Will have re-ordered, so start
1299 * from beginning again */
1301 } else if (type >= LALR) {
1302 struct symset ss = set_find(g, is->items.data[pos]);
1303 struct symset tmp = INIT_SYMSET;
1305 symset_union(&tmp, &ss);
1306 if (symset_union(&tmp, &LA)) {
1307 is->items.data[pos] = save_set(g, tmp);
1315 For each symbol we found (and placed in `done`) we collect all the items for
1316 which this symbol is next, and create a new itemset with "DOT" advanced over
1317 the symbol. This is then added to the collection of itemsets (or merged
1318 with a pre-existing itemset).
1320 ###### derive itemsets
1321 // Now we have a completed itemset, so we need to
1322 // compute all the 'next' itemsets and create them
1323 // if they don't exist.
1324 for (i = 0; i < done.cnt; i++) {
1326 unsigned short state;
1327 int starts_line = 0;
1328 struct symbol *sym = g->symtab[done.syms[i]];
1329 struct symset newitemset = INIT_SYMSET;
1331 newitemset = INIT_DATASET;
1334 (sym->nullable && is->starts_line))
1336 for (j = 0; j < is->items.cnt; j++) {
1337 int itm = is->items.syms[j];
1338 int p = item_prod(itm);
1339 int bp = item_index(itm);
1340 struct production *pr = g->productions[p];
1341 unsigned short la = 0;
1344 if (bp == pr->body_size)
1346 if (pr->body[bp] != sym)
1349 la = is->items.data[j];
1350 pos = symset_find(&newitemset, pr->head->num);
1352 symset_add(&newitemset, item_num(p, bp+1), la);
1353 else if (type >= LALR) {
1354 // Need to merge la set.
1355 int la2 = newitemset.data[pos];
1357 struct symset ss = set_find(g, la2);
1358 struct symset LA = INIT_SYMSET;
1359 symset_union(&LA, &ss);
1360 ss = set_find(g, la);
1361 if (symset_union(&LA, &ss))
1362 newitemset.data[pos] = save_set(g, LA);
1368 state = add_itemset(g, newitemset, type, starts_line);
1369 if (symset_find(&is->go_to, done.syms[i]) < 0)
1370 symset_add(&is->go_to, done.syms[i], state);
1373 All that is left is to crate the initial itemset from production zero, and
1374 with `TK_eof` as the LA set.
1377 static void build_itemsets(struct grammar *g, enum grammar_type type)
1379 struct symset first = INIT_SYMSET;
1382 unsigned short la = 0;
1384 // LA set just has eof
1385 struct symset eof = INIT_SYMSET;
1386 symset_add(&eof, TK_eof, 0);
1387 la = save_set(g, eof);
1388 first = INIT_DATASET;
1390 // production 0, offset 0 (with no data)
1391 symset_add(&first, item_num(0, 0), la);
1392 add_itemset(g, first, type, g->productions[0]->body[0]->can_eol);
1393 for (again = 0, is = g->items;
1395 is = is->next ?: again ? (again = 0, g->items) : NULL) {
1397 struct symset done = INIT_SYMSET;
1408 ### Completing the analysis.
1410 The exact process of analysis depends on which level was requested. For
1411 `LR0` and `LR05` we don't need first and follow sets at all. For `LALR` and
1412 `LR1` we need first, but not follow. For `SLR` we need both.
1414 We don't build the "action" tables that you might expect as the parser
1415 doesn't use them. They will be calculated to some extent if needed for
1418 Once we have built everything we allocate arrays for the two lists:
1419 symbols and itemsets. This allows more efficient access during reporting.
1420 The symbols are grouped as terminals and non-terminals and we record the
1421 changeover point in `first_nonterm`.
1423 ###### grammar fields
1424 struct symbol **symtab;
1425 struct itemset **statetab;
1428 ###### grammar_analyse
1430 static void grammar_analyse(struct grammar *g, enum grammar_type type)
1434 int snum = TK_reserved;
1435 for (s = g->syms; s; s = s->next)
1436 if (s->num < 0 && s->type == Terminal) {
1440 g->first_nonterm = snum;
1441 for (s = g->syms; s; s = s->next)
1447 g->symtab = calloc(g->num_syms, sizeof(g->symtab[0]));
1448 for (s = g->syms; s; s = s->next)
1449 g->symtab[s->num] = s;
1459 build_itemsets(g, type);
1461 g->statetab = calloc(g->states, sizeof(g->statetab[0]));
1462 for (is = g->items; is ; is = is->next)
1463 g->statetab[is->state] = is;
1466 ## Reporting on the Grammar
1468 The purpose of the report is to give the grammar developer insight into
1469 how the grammar parser will work. It is basically a structured dump of
1470 all the tables that have been generated, plus a description of any conflicts.
1472 ###### grammar_report
1473 static int grammar_report(struct grammar *g, enum grammar_type type)
1479 return report_conflicts(g, type);
1482 Firstly we have the complete list of symbols, together with the
1483 "FIRST" set if that was generated. We add a mark to each symbol to
1484 show if it can end in a newline (`>`), or if it is nullable (`.`).
1488 static void report_symbols(struct grammar *g)
1492 printf("SYMBOLS + FIRST:\n");
1494 printf("SYMBOLS:\n");
1496 for (n = 0; n < g->num_syms; n++) {
1497 struct symbol *s = g->symtab[n];
1501 printf(" %c%c%3d%c: ",
1502 s->nullable ? '.':' ',
1503 s->can_eol ? '>':' ',
1504 s->num, symtypes[s->type]);
1507 printf(" (%d%s)", s->precedence,
1508 assoc_names[s->assoc]);
1510 if (g->first && s->type == Nonterminal) {
1513 for (j = 0; j < g->first[n].cnt; j++) {
1516 prtxt(g->symtab[g->first[n].syms[j]]->name);
1523 Then we have the follow sets if they were computed.
1525 static void report_follow(struct grammar *g)
1528 printf("FOLLOW:\n");
1529 for (n = 0; n < g->num_syms; n++)
1530 if (g->follow[n].cnt) {
1534 prtxt(g->symtab[n]->name);
1535 for (j = 0; j < g->follow[n].cnt; j++) {
1538 prtxt(g->symtab[g->follow[n].syms[j]]->name);
1544 And finally the item sets. These include the GO TO tables and, for
1545 LALR and LR1, the LA set for each item. Lots of stuff, so we break
1546 it up a bit. First the items, with production number and associativity.
1548 static void report_item(struct grammar *g, int itm)
1550 int p = item_prod(itm);
1551 int dot = item_index(itm);
1552 struct production *pr = g->productions[p];
1556 prtxt(pr->head->name);
1558 for (i = 0; i < pr->body_size; i++) {
1559 printf(" %s", (dot == i ? ". ": ""));
1560 prtxt(pr->body[i]->name);
1562 if (dot == pr->body_size)
1566 printf(" (%d%s)", pr->precedence,
1567 assoc_names[pr->assoc]);
1571 The LA sets which are (possibly) reported with each item:
1573 static void report_la(struct grammar *g, int lanum)
1575 struct symset la = set_find(g, lanum);
1579 printf(" LOOK AHEAD(%d)", lanum);
1580 for (i = 0; i < la.cnt; i++) {
1583 prtxt(g->symtab[la.syms[i]]->name);
1588 Then the go to sets:
1591 static void report_goto(struct grammar *g, struct symset gt)
1596 for (i = 0; i < gt.cnt; i++) {
1598 prtxt(g->symtab[gt.syms[i]]->name);
1599 printf(" -> %d\n", gt.data[i]);
1603 Now we can report all the item sets complete with items, LA sets, and GO TO.
1605 static void report_itemsets(struct grammar *g)
1608 printf("ITEM SETS(%d)\n", g->states);
1609 for (s = 0; s < g->states; s++) {
1611 struct itemset *is = g->statetab[s];
1612 printf(" Itemset %d:%s\n", s, is->starts_line?" (startsline)":"");
1613 for (j = 0; j < is->items.cnt; j++) {
1614 report_item(g, is->items.syms[j]);
1615 if (is->items.data != NO_DATA)
1616 report_la(g, is->items.data[j]);
1618 report_goto(g, is->go_to);
1622 ### Reporting conflicts
1624 Conflict detection varies a lot among different analysis levels. However
1625 LR0 and LR0.5 are quite similar - having no follow sets, and SLR, LALR and
1626 LR1 are also similar as they have FOLLOW or LA sets.
1630 ## conflict functions
1632 static int report_conflicts(struct grammar *g, enum grammar_type type)
1635 printf("Conflicts:\n");
1637 cnt = conflicts_lr0(g, type);
1639 cnt = conflicts_slr(g, type);
1641 printf(" - no conflicts\n");
1645 LR0 conflicts are any state which have both a reducible item and
1646 a shiftable item, or two reducible items.
1648 LR05 conflicts only occurs if two possibly reductions exist,
1649 as shifts always over-ride reductions.
1651 ###### conflict functions
1652 static int conflicts_lr0(struct grammar *g, enum grammar_type type)
1656 for (i = 0; i < g->states; i++) {
1657 struct itemset *is = g->statetab[i];
1658 int last_reduce = -1;
1659 int prev_reduce = -1;
1660 int last_shift = -1;
1664 for (j = 0; j < is->items.cnt; j++) {
1665 int itm = is->items.syms[j];
1666 int p = item_prod(itm);
1667 int bp = item_index(itm);
1668 struct production *pr = g->productions[p];
1670 if (bp == pr->body_size) {
1671 prev_reduce = last_reduce;
1675 if (pr->body[bp]->type == Terminal)
1678 if (type == LR0 && last_reduce >= 0 && last_shift >= 0) {
1679 printf(" State %d has both SHIFT and REDUCE:\n", i);
1680 report_item(g, is->items.syms[last_shift]);
1681 report_item(g, is->items.syms[last_reduce]);
1684 if (prev_reduce >= 0) {
1685 printf(" State %d has 2 (or more) reducible items\n", i);
1686 report_item(g, is->items.syms[prev_reduce]);
1687 report_item(g, is->items.syms[last_reduce]);
1694 SLR, LALR, and LR1 conflicts happen if two reducible items have over-lapping
1695 look ahead, or if a symbol in a look-ahead can be shifted. They differ only
1696 in the source of the look ahead set.
1698 We build two datasets to reflect the "action" table: one which maps
1699 terminals to items where that terminal could be shifted and another
1700 which maps terminals to items that could be reduced when the terminal
1701 is in look-ahead. We report when we get conflicts between the two.
1703 static int conflicts_slr(struct grammar *g, enum grammar_type type)
1708 for (i = 0; i < g->states; i++) {
1709 struct itemset *is = g->statetab[i];
1710 struct symset shifts = INIT_DATASET;
1711 struct symset reduce = INIT_DATASET;
1715 /* First collect the shifts */
1716 for (j = 0; j < is->items.cnt; j++) {
1717 unsigned short itm = is->items.syms[j];
1718 int p = item_prod(itm);
1719 int bp = item_index(itm);
1720 struct production *pr = g->productions[p];
1722 if (bp < pr->body_size &&
1723 pr->body[bp]->type == Terminal) {
1725 int sym = pr->body[bp]->num;
1726 if (symset_find(&shifts, sym) < 0)
1727 symset_add(&shifts, sym, itm);
1730 /* Now look for reduction and conflicts */
1731 for (j = 0; j < is->items.cnt; j++) {
1732 unsigned short itm = is->items.syms[j];
1733 int p = item_prod(itm);
1734 int bp = item_index(itm);
1735 struct production *pr = g->productions[p];
1737 if (bp < pr->body_size)
1742 la = g->follow[pr->head->num];
1744 la = set_find(g, is->items.data[j]);
1746 for (k = 0; k < la.cnt; k++) {
1747 int pos = symset_find(&shifts, la.syms[k]);
1749 printf(" State %d has SHIFT/REDUCE conflict on ", i);
1750 prtxt(g->symtab[la.syms[k]]->name);
1752 report_item(g, shifts.data[pos]);
1753 report_item(g, itm);
1756 pos = symset_find(&reduce, la.syms[k]);
1758 symset_add(&reduce, la.syms[k], itm);
1761 printf(" State %d has REDUCE/REDUCE conflict on ", i);
1762 prtxt(g->symtab[la.syms[k]]->name);
1764 report_item(g, itm);
1765 report_item(g, reduce.data[pos]);
1769 symset_free(shifts);
1770 symset_free(reduce);
1776 ## Generating the parser
1778 The exported part of the parser is the `parse_XX` function, where the name
1779 `XX` is based on the name of the parser files.
1781 This takes a `code_node`, a partially initialized `token_config`, and an
1782 optional `FILE` to send tracing to. The `token_config` gets the list of
1783 known words added and then is used with the `code_node` to initialize the
1786 `parse_XX` then calls the library function `parser_run` to actually complete
1787 the parse. This needs the `states` table and function to call the various
1788 pieces of code provided in the grammar file, so they are generated first.
1790 ###### parser_generate
1792 static void gen_parser(FILE *f, struct grammar *g, char *file, char *name)
1798 gen_reduce(f, g, file);
1801 fprintf(f, "#line 0 \"gen_parser\"\n");
1802 fprintf(f, "void *parse_%s(struct code_node *code, struct token_config *config, FILE *trace)\n",
1805 fprintf(f, "\tstruct token_state *tokens;\n");
1806 fprintf(f, "\tconfig->words_marks = known;\n");
1807 fprintf(f, "\tconfig->known_count = sizeof(known)/sizeof(known[0]);\n");
1808 fprintf(f, "\tconfig->ignored |= (1 << TK_line_comment) | (1 << TK_block_comment);\n");
1809 fprintf(f, "\ttokens = token_open(code, config);\n");
1810 fprintf(f, "\tvoid *rv = parser_run(tokens, states, do_reduce, do_free, trace, non_term, config);\n");
1811 fprintf(f, "\ttoken_close(tokens);\n");
1812 fprintf(f, "\treturn rv;\n");
1813 fprintf(f, "}\n\n");
1816 ### Known words table
1818 The known words table is simply an array of terminal symbols.
1819 The table of nonterminals used for tracing is a similar array.
1823 static void gen_known(FILE *f, struct grammar *g)
1826 fprintf(f, "#line 0 \"gen_known\"\n");
1827 fprintf(f, "static const char *known[] = {\n");
1828 for (i = TK_reserved;
1829 i < g->num_syms && g->symtab[i]->type == Terminal;
1831 fprintf(f, "\t\"%.*s\",\n", g->symtab[i]->name.len,
1832 g->symtab[i]->name.txt);
1833 fprintf(f, "};\n\n");
1836 static void gen_non_term(FILE *f, struct grammar *g)
1839 fprintf(f, "#line 0 \"gen_non_term\"\n");
1840 fprintf(f, "static const char *non_term[] = {\n");
1841 for (i = TK_reserved;
1844 if (g->symtab[i]->type == Nonterminal)
1845 fprintf(f, "\t\"%.*s\",\n", g->symtab[i]->name.len,
1846 g->symtab[i]->name.txt);
1847 fprintf(f, "};\n\n");
1850 ### States and the goto tables.
1852 For each state we record the goto table, the reducible production if
1853 there is one, or a symbol to shift for error recovery.
1854 Some of the details of the reducible production are stored in the
1855 `do_reduce` function to come later. Here we store the production number,
1856 the body size (useful for stack management) and the resulting symbol (useful
1857 for knowing how to free data later).
1859 The go to table is stored in a simple array of `sym` and corresponding
1862 ###### exported types
1870 const struct lookup * go_to;
1881 static void gen_goto(FILE *f, struct grammar *g)
1884 fprintf(f, "#line 0 \"gen_goto\"\n");
1885 for (i = 0; i < g->states; i++) {
1887 fprintf(f, "static const struct lookup goto_%d[] = {\n",
1889 struct symset gt = g->statetab[i]->go_to;
1890 for (j = 0; j < gt.cnt; j++)
1891 fprintf(f, "\t{ %d, %d },\n",
1892 gt.syms[j], gt.data[j]);
1899 static void gen_states(FILE *f, struct grammar *g)
1902 fprintf(f, "#line 0 \"gen_states\"\n");
1903 fprintf(f, "static const struct state states[] = {\n");
1904 for (i = 0; i < g->states; i++) {
1905 struct itemset *is = g->statetab[i];
1906 int j, prod = -1, prod_len;
1908 int shift_len = 0, shift_remain = 0;
1909 for (j = 0; j < is->items.cnt; j++) {
1910 int itm = is->items.syms[j];
1911 int p = item_prod(itm);
1912 int bp = item_index(itm);
1913 struct production *pr = g->productions[p];
1915 if (bp < pr->body_size) {
1916 if (shift_sym < 0 ||
1917 (shift_len == bp && shift_remain > pr->body_size - bp)) {
1918 shift_sym = pr->body[bp]->num;
1920 shift_remain = pr->body_size - bp;
1924 /* This is what we reduce */
1925 if (prod < 0 || prod_len < pr->body_size) {
1927 prod_len = pr->body_size;
1932 fprintf(f, "\t[%d] = { %d, goto_%d, %d, %d, %d, 0, %d },\n",
1933 i, is->go_to.cnt, i, prod,
1934 g->productions[prod]->body_size,
1935 g->productions[prod]->head->num,
1938 fprintf(f, "\t[%d] = { %d, goto_%d, -1, -1, -1, %d, %d },\n",
1939 i, is->go_to.cnt, i, shift_sym,
1942 fprintf(f, "};\n\n");
1945 ### The `do_reduce` function and the code
1947 When the parser engine decides to reduce a production, it calls `do_reduce`.
1948 This has two functions.
1950 Firstly, if a non-NULL `trace` file is passed, it prints out details of the
1951 production being reduced. Secondly it runs the code that was included with
1952 the production if any.
1954 This code needs to be able to store data somewhere. Rather than requiring
1955 `do_reduce` to `malloc` that "somewhere", we pass in a large buffer and have
1956 `do_reduce` return the size to be saved.
1958 In order for the code to access "global" context, we pass in the
1959 "config" pointer that was passed to parser function. If the `struct
1960 token_config` is embedded in some larger structure, the reducing code
1961 can access the larger structure using pointer manipulation.
1963 The code fragment requires translation when written out. Any `$N` needs to
1964 be converted to a reference either to that buffer (if `$0`) or to the
1965 structure returned by a previous reduction. These pointers need to be cast
1966 to the appropriate type for each access. All this is handling in
1969 `gen_code` also allows symbol references to contain a '`<`' as in '`$<2`'.
1970 This applied only to symbols with references (or pointers), not those with structures.
1971 The `<` implies that the reference it being moved out, so the object will not be
1972 automatically freed. This is equivalent to assigning `NULL` to the pointer.
1976 static void gen_code(struct production *p, FILE *f, struct grammar *g)
1979 char *used = calloc(1, p->body_size);
1982 fprintf(f, "\t\t\t");
1983 for (c = p->code.txt; c < p->code.txt + p->code.len; c++) {
1997 if (*c < '0' || *c > '9') {
2004 while (c[1] >= '0' && c[1] <= '9') {
2006 n = n * 10 + *c - '0';
2009 fprintf(f, "(*(struct %.*s*%s)ret)",
2010 p->head->struct_name.len,
2011 p->head->struct_name.txt,
2012 p->head->isref ? "*":"");
2013 else if (n > p->body_size)
2014 fprintf(f, "$%d", n);
2015 else if (p->body[n-1]->type == Terminal)
2016 fprintf(f, "(*(struct token *)body[%d])",
2018 else if (p->body[n-1]->struct_name.txt == NULL)
2019 fprintf(f, "$%d", n);
2021 fprintf(f, "(*(struct %.*s*%s)body[%d])",
2022 p->body[n-1]->struct_name.len,
2023 p->body[n-1]->struct_name.txt,
2024 p->body[n-1]->isref ? "*":"", n-1);
2029 for (i = 0; i < p->body_size; i++) {
2030 if (p->body[i]->struct_name.txt &&
2031 p->body[i]->isref &&
2033 // assume this has been copied out
2034 fprintf(f, "\t\t*(void**)body[%d] = NULL;\n", i);
2041 static void gen_reduce(FILE *f, struct grammar *g, char *file)
2044 fprintf(f, "#line 0 \"gen_reduce\"\n");
2045 fprintf(f, "static int do_reduce(int prod, void **body, struct token_config *config, void *ret)\n");
2047 fprintf(f, "\tint ret_size = 0;\n");
2049 fprintf(f, "\tswitch(prod) {\n");
2050 for (i = 0; i < g->production_count; i++) {
2051 struct production *p = g->productions[i];
2052 fprintf(f, "\tcase %d:\n", i);
2057 if (p->head->struct_name.txt)
2058 fprintf(f, "\t\tret_size = sizeof(struct %.*s%s);\n",
2059 p->head->struct_name.len,
2060 p->head->struct_name.txt,
2061 p->head->isref ? "*":"");
2063 fprintf(f, "\t\tbreak;\n");
2065 fprintf(f, "\t}\n\treturn ret_size;\n}\n\n");
2070 As each non-terminal can potentially cause a different type of data
2071 structure to be allocated and filled in, we need to be able to free it when
2074 It is particularly important to have fine control over freeing during error
2075 recovery where individual stack frames might need to be freed.
2077 For this, the grammar author is required to defined a `free_XX` function for
2078 each structure that is used by a non-terminal. `do_free` will call whichever
2079 is appropriate given a symbol number, and will call `free` (as is
2080 appropriate for tokens on any terminal symbol.
2084 static void gen_free(FILE *f, struct grammar *g)
2088 fprintf(f, "#line 0 \"gen_free\"\n");
2089 fprintf(f, "static void do_free(short sym, void *asn)\n");
2091 fprintf(f, "\tif (!asn) return;\n");
2092 fprintf(f, "\tif (sym < %d) {\n", g->first_nonterm);
2093 fprintf(f, "\t\tfree(asn);\n\t\treturn;\n\t}\n");
2094 fprintf(f, "\tswitch(sym) {\n");
2096 for (i = 0; i < g->num_syms; i++) {
2097 struct symbol *s = g->symtab[i];
2099 s->type != Nonterminal ||
2100 s->struct_name.txt == NULL)
2103 fprintf(f, "\tcase %d:\n", s->num);
2105 fprintf(f, "\t\tfree_%.*s(*(void**)asn);\n",
2107 s->struct_name.txt);
2108 fprintf(f, "\t\tfree(asn);\n");
2110 fprintf(f, "\t\tfree_%.*s(asn);\n",
2112 s->struct_name.txt);
2113 fprintf(f, "\t\tbreak;\n");
2115 fprintf(f, "\t}\n}\n\n");
2118 ## The main routine.
2120 There are three key parts to the "main" function of parsergen: processing
2121 the arguments, loading the grammar file, and dealing with the grammar.
2123 ### Argument processing.
2125 Command line options allow the selection of analysis mode, name of output
2126 file, and whether or not a report should be generated. By default we create
2127 a report only if no code output was requested.
2129 The `parse_XX` function name uses the basename of the output file which
2130 should not have a suffix (`.c`). `.c` is added to the given name for the
2131 code, and `.h` is added for the header (if header text is specifed in the
2138 static const struct option long_options[] = {
2139 { "LR0", 0, NULL, '0' },
2140 { "LR05", 0, NULL, '5' },
2141 { "SLR", 0, NULL, 'S' },
2142 { "LALR", 0, NULL, 'L' },
2143 { "LR1", 0, NULL, '1' },
2144 { "tag", 1, NULL, 't' },
2145 { "report", 0, NULL, 'R' },
2146 { "output", 1, NULL, 'o' },
2147 { NULL, 0, NULL, 0 }
2149 const char *options = "05SL1t:Ro:";
2151 ###### process arguments
2153 char *outfile = NULL;
2158 enum grammar_type type = LR05;
2159 while ((opt = getopt_long(argc, argv, options,
2160 long_options, NULL)) != -1) {
2175 outfile = optarg; break;
2177 tag = optarg; break;
2179 fprintf(stderr, "Usage: parsergen ...\n");
2184 infile = argv[optind++];
2186 fprintf(stderr, "No input file given\n");
2189 if (outfile && report == 1)
2192 if (name && strchr(name, '/'))
2193 name = strrchr(name, '/')+1;
2195 if (optind < argc) {
2196 fprintf(stderr, "Excess command line arguments\n");
2200 ### Loading the grammar file
2202 To be able to run `mdcode` and `scanner` on the grammar we need to memory
2205 One we have extracted the code (with `mdcode`) we expect to find three
2206 sections: header, code, and grammar. Anything else that is not
2207 excluded by the `--tag` option is an error.
2209 "header" and "code" are optional, though it is hard to build a working
2210 parser with neither. "grammar" must be provided.
2214 #include <sys/mman.h>
2219 static void pr_err(char *msg)
2222 fprintf(stderr, "%s\n", msg);
2226 struct section *table;
2230 fd = open(infile, O_RDONLY);
2232 fprintf(stderr, "parsergen: cannot open %s: %s\n",
2233 infile, strerror(errno));
2236 len = lseek(fd, 0, 2);
2237 file = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0);
2238 table = code_extract(file, file+len, pr_err);
2240 struct code_node *hdr = NULL;
2241 struct code_node *code = NULL;
2242 struct code_node *gram = NULL;
2243 for (s = table; s; s = s->next) {
2244 struct text sec = s->section;
2245 if (tag && !strip_tag(&sec, tag))
2247 if (text_is(sec, "header"))
2249 else if (text_is(sec, "code"))
2251 else if (text_is(sec, "grammar"))
2254 fprintf(stderr, "Unknown content section: %.*s\n",
2255 s->section.len, s->section.txt);
2260 ### Processing the input
2262 As we need to append an extention to a filename and then open it for
2263 writing, and we need to do this twice, it helps to have a separate function.
2267 static FILE *open_ext(char *base, char *ext)
2269 char *fn = malloc(strlen(base) + strlen(ext) + 1);
2271 strcat(strcpy(fn, base), ext);
2277 If we can read the grammar, then we analyse and optionally report on it. If
2278 the report finds conflicts we will exit with an error status.
2280 ###### process input
2281 struct grammar *g = NULL;
2283 fprintf(stderr, "No grammar section provided\n");
2286 g = grammar_read(gram);
2288 fprintf(stderr, "Failure to parse grammar\n");
2293 grammar_analyse(g, type);
2295 if (grammar_report(g, type))
2299 If a headers section is defined, we write it out and include a declaration
2300 for the `parse_XX` function so it can be used from separate code.
2302 if (rv == 0 && hdr && outfile) {
2303 FILE *f = open_ext(outfile, ".h");
2305 code_node_print(f, hdr, infile);
2306 fprintf(f, "void *parse_%s(struct code_node *code, struct token_config *config, FILE *trace);\n",
2310 fprintf(stderr, "Cannot create %s.h\n",
2316 And if all goes well and an output file was provided, we create the `.c`
2317 file with the code section (if any) and the parser tables and function.
2319 if (rv == 0 && outfile) {
2320 FILE *f = open_ext(outfile, ".c");
2323 code_node_print(f, code, infile);
2324 gen_parser(f, g, infile, name);
2327 fprintf(stderr, "Cannot create %s.c\n",
2333 And that about wraps it up. We need to set the locale so that UTF-8 is
2334 recognised properly, and link with `libicuuc` as `libmdcode` requires that.
2336 ###### File: parsergen.mk
2337 parsergen : parsergen.o libscanner.o libmdcode.o
2338 $(CC) $(CFLAGS) -o parsergen parsergen.o libscanner.o libmdcode.o -licuuc
2345 int main(int argc, char *argv[])
2350 setlocale(LC_ALL,"");
2352 ## process arguments
2359 ## The SHIFT/REDUCE parser
2361 Having analysed the grammar and generated all the tables, we only need the
2362 shift/reduce engine to bring it all together.
2364 ### Goto table lookup
2366 The parser generator has nicely provided us with goto tables sorted by
2367 symbol number. We need a binary search function to find a symbol in the
2370 ###### parser functions
2372 static int search(const struct state *l, int sym)
2375 int hi = l->go_to_cnt;
2379 while (lo + 1 < hi) {
2380 int mid = (lo + hi) / 2;
2381 if (l->go_to[mid].sym <= sym)
2386 if (l->go_to[lo].sym == sym)
2387 return l->go_to[lo].state;
2392 ### The state stack.
2394 The core data structure for the parser is the stack. This tracks all the
2395 symbols that have been recognised or partially recognised.
2397 The stack usually won't grow very large - maybe a few tens of entries. So
2398 we dynamically resize and array as required but never bother to shrink it
2401 We keep the stack as two separate allocations. One, `asn_stack` stores the
2402 "abstract syntax nodes" which are created by each reduction. When we call
2403 `do_reduce` we need to pass an array of the `asn`s of the body of the
2404 production, and by keeping a separate `asn` stack, we can just pass a
2405 pointer into this stack.
2407 The other allocation stores all other stack fields of which there are six.
2408 The `state` is the most important one and guides the parsing process. The
2409 `sym` is nearly unnecessary. However when we want to free entries from the
2410 `asn_stack`, it helps to know what type they are so we can call the right
2411 freeing function. The symbol leads us to the right free function through
2414 The `indents` count and the `starts_indented` flag track the line
2415 indents in the symbol. These are used to allow indent information to
2416 guide parsing and error recovery.
2418 `newline_permitted` keeps track of whether newlines should be ignored
2419 or not, and `starts_line` records if this state stated on a newline.
2421 As well as the stack of frames we have a `next` frame which is
2422 assembled from the incoming token and other information prior to
2423 pushing it onto the stack.
2425 ###### parser functions
2431 short starts_indented;
2433 short newline_permitted;
2442 Two operations are needed on the stack - shift (which is like push) and pop.
2444 Shift applies not only to terminals but also to non-terminals. When we
2445 reduce a production we will pop off entries corresponding to the body
2446 symbols, then push on an item for the head of the production. This last is
2447 exactly the same process as shifting in a terminal so we use the same
2450 To simplify other code we arrange for `shift` to fail if there is no `goto`
2451 state for the symbol. This is useful in basic parsing due to our design
2452 that we shift when we can, and reduce when we cannot. So the `shift`
2453 function reports if it could.
2455 So `shift` finds the next state. If that succeed it extends the allocations
2456 if needed and pushes all the information onto the stacks.
2458 ###### parser functions
2460 static int shift(struct parser *p,
2462 const struct state states[])
2464 // Push an entry onto the stack
2465 int newstate = search(&states[p->next.state], p->next.sym);
2468 if (p->tos >= p->stack_size) {
2469 p->stack_size += 10;
2470 p->stack = realloc(p->stack, p->stack_size
2471 * sizeof(p->stack[0]));
2472 p->asn_stack = realloc(p->asn_stack, p->stack_size
2473 * sizeof(p->asn_stack[0]));
2475 p->stack[p->tos] = p->next;
2476 p->asn_stack[p->tos] = asn;
2478 p->next.state = newstate;
2479 p->next.indents = 0;
2480 p->next.starts_indented = 0;
2481 // if new state doesn't start a line, we inherit newline_permitted status
2482 if (states[newstate].starts_line)
2483 p->next.newline_permitted = 1;
2487 `pop` simply moves the top of stack (`tos`) back down the required amount
2488 and frees any `asn` entries that need to be freed. It is called _after_ we
2489 reduce a production, just before we `shift` the nonterminal in.
2491 ###### parser functions
2493 static void pop(struct parser *p, int num,
2494 void(*do_free)(short sym, void *asn))
2498 for (i = 0; i < num; i++) {
2499 p->next.indents += p->stack[p->tos+i].indents;
2500 do_free(p->stack[p->tos+i].sym,
2501 p->asn_stack[p->tos+i]);
2505 p->next.state = p->stack[p->tos].state;
2506 p->next.starts_indented = p->stack[p->tos].starts_indented;
2507 p->next.newline_permitted = p->stack[p->tos].newline_permitted;
2508 if (p->next.indents > p->next.starts_indented)
2509 p->next.newline_permitted = 0;
2513 ### Memory allocation
2515 The `scanner` returns tokens in a local variable - we want them in allocated
2516 memory so they can live in the `asn_stack`. Similarly the `asn` produced by
2517 a reduce is in a large buffer. Both of these require some allocation and
2518 copying, hence `memdup` and `tokcopy`.
2520 ###### parser includes
2523 ###### parser functions
2525 void *memdup(void *m, int len)
2531 memcpy(ret, m, len);
2535 static struct token *tok_copy(struct token tk)
2537 struct token *new = malloc(sizeof(*new));
2542 ### The heart of the parser.
2544 Now we have the parser. If we can shift, we do. If not and we can reduce
2545 we do. If the production we reduced was production zero, then we have
2546 accepted the input and can finish.
2548 We return whatever `asn` was returned by reducing production zero.
2550 If we can neither shift nor reduce we have an error to handle. We pop
2551 single entries off the stack until we can shift the `TK_error` symbol, then
2552 drop input tokens until we find one we can shift into the new error state.
2554 When we find `TK_in` and `TK_out` tokens which report indents we need
2555 to handle them directly as the grammar cannot express what we want to
2558 `TK_in` tokens are easy: we simply update the `next` stack frame to
2559 record how many indents there are and that the next token started with
2562 `TK_out` tokens must either be counted off against any pending indent,
2563 or must force reductions until there is a pending indent which isn't
2564 at the start of a production.
2566 `TK_newline` tokens are ignored precisely if there has been an indent
2567 since the last state which could have been at the start of a line.
2569 ###### parser includes
2572 void *parser_run(struct token_state *tokens,
2573 const struct state states[],
2574 int (*do_reduce)(int, void**, struct token_config*, void*),
2575 void (*do_free)(short, void*),
2576 FILE *trace, const char *non_term[],
2577 struct token_config *config)
2579 struct parser p = { 0 };
2580 struct token *tk = NULL;
2584 p.next.newline_permitted = states[0].starts_line;
2586 struct token *err_tk;
2588 tk = tok_copy(token_next(tokens));
2589 p.next.sym = tk->num;
2591 parser_trace(trace, &p, tk, states, non_term, config->known_count);
2593 if (p.next.sym == TK_in) {
2594 p.next.starts_indented = 1;
2600 if (p.next.sym == TK_out) {
2601 if (p.stack[p.tos-1].indents > p.stack[p.tos-1].starts_indented ||
2602 (p.stack[p.tos-1].indents == 1 &&
2603 states[p.next.state].reduce_size != 1)) {
2604 p.stack[p.tos-1].indents -= 1;
2605 if (p.stack[p.tos-1].indents == p.stack[p.tos-1].starts_indented) {
2606 // no internal indent any more, reassess 'newline_permitted'
2607 if (states[p.stack[p.tos-1].state].starts_line)
2608 p.stack[p.tos-1].newline_permitted = 1;
2610 p.stack[p.tos-1].newline_permitted = p.stack[p.tos-2].newline_permitted;
2616 // fall through and force a REDUCE (as 'shift'
2619 if (p.next.sym == TK_newline) {
2620 if (!p.tos || ! p.stack[p.tos-1].newline_permitted) {
2626 if (shift(&p, tk, states)) {
2630 if (states[p.next.state].reduce_prod >= 0) {
2633 int prod = states[p.next.state].reduce_prod;
2634 int size = states[p.next.state].reduce_size;
2636 static char buf[16*1024];
2637 p.next.sym = states[p.next.state].reduce_sym;
2639 body = p.asn_stack +
2640 (p.tos - states[p.next.state].reduce_size);
2642 bufsize = do_reduce(prod, body, config, buf);
2644 pop(&p, size, do_free);
2645 res = memdup(buf, bufsize);
2646 memset(buf, 0, bufsize);
2647 if (!shift(&p, res, states)) {
2648 if (prod != 0) abort();
2654 if (tk->num == TK_out) {
2655 // Indent problem - synthesise tokens to get us
2657 fprintf(stderr, "Synthesize %d to handle indent problem\n", states[p.next.state].shift_sym);
2658 p.next.sym = states[p.next.state].shift_sym;
2659 shift(&p, tok_copy(*tk), states);
2660 // FIXME need to report this error somehow
2663 /* Error. We walk up the stack until we
2664 * find a state which will accept TK_error.
2665 * We then shift in TK_error and see what state
2666 * that takes us too.
2667 * Then we discard input tokens until
2668 * we find one that is acceptable.
2671 err_tk = tok_copy(*tk);
2672 p.next.sym = TK_error;
2673 while (shift(&p, err_tk, states) == 0
2675 // discard this state
2676 pop(&p, 1, do_free);
2679 // no state accepted TK_error
2682 while (search(&states[p.next.state], tk->num) < 0 &&
2683 tk->num != TK_eof) {
2685 tk = tok_copy(token_next(tokens));
2686 if (tk->num == TK_in)
2687 p.next.indents += 1;
2688 if (tk->num == TK_out) {
2689 if (p.next.indents == 0)
2691 p.next.indents -= 1;
2694 if (p.tos == 0 && tk->num == TK_eof)
2698 pop(&p, p.tos, do_free);
2704 ###### exported functions
2705 void *parser_run(struct token_state *tokens,
2706 const struct state states[],
2707 int (*do_reduce)(int, void**, struct token_config*, void*),
2708 void (*do_free)(short, void*),
2709 FILE *trace, const char *non_term[],
2710 struct token_config *config);
2714 Being able to visualize the parser in action can be invaluable when
2715 debugging the parser code, or trying to understand how the parse of a
2716 particular grammar progresses. The stack contains all the important
2717 state, so just printing out the stack every time around the parse loop
2718 can make it possible to see exactly what is happening.
2720 This doesn't explicitly show each SHIFT and REDUCE action. However they
2721 are easily deduced from the change between consecutive lines, and the
2722 details of each state can be found by cross referencing the states list
2723 in the stack with the "report" that parsergen can generate.
2725 For terminal symbols, we just dump the token. For non-terminals we
2726 print the name of the symbol. The look ahead token is reported at the
2727 end inside square brackets.
2729 ###### parser functions
2731 static char *reserved_words[] = {
2732 [TK_error] = "ERROR",
2735 [TK_newline] = "NEWLINE",
2738 static void parser_trace_state(FILE *trace, struct frame *f, const struct state states[])
2740 fprintf(trace, "(%d", f->state);
2742 fprintf(trace, "%c%d", f->starts_indented?':':'.',
2744 if (states[f->state].starts_line)
2745 fprintf(trace, "s");
2746 if (f->newline_permitted)
2747 fprintf(trace, "n");
2748 fprintf(trace, ") ");
2751 void parser_trace(FILE *trace, struct parser *p,
2752 struct token *tk, const struct state states[],
2753 const char *non_term[], int knowns)
2756 for (i = 0; i < p->tos; i++) {
2757 int sym = p->stack[i].sym;
2758 parser_trace_state(trace, &p->stack[i], states);
2759 if (sym < TK_reserved &&
2760 reserved_words[sym] != NULL)
2761 fputs(reserved_words[sym], trace);
2762 else if (sym < TK_reserved + knowns) {
2763 struct token *t = p->asn_stack[i];
2764 text_dump(trace, t->txt, 20);
2766 fputs(non_term[sym - TK_reserved - knowns],
2770 parser_trace_state(trace, &p->next, states);
2771 fprintf(trace, " [");
2772 if (tk->num < TK_reserved &&
2773 reserved_words[tk->num] != NULL)
2774 fputs(reserved_words[tk->num], trace);
2776 text_dump(trace, tk->txt, 20);
2777 fputs("]\n", trace);
2782 The obvious example for a parser is a calculator.
2784 As `scanner` provides number parsing function using `libgmp` is it not much
2785 work to perform arbitrary rational number calculations.
2787 This calculator takes one expression, or an equality test per line. The
2788 results are printed and if any equality test fails, the program exits with
2791 ###### File: parsergen.mk
2792 calc.c calc.h : parsergen parsergen.mdc
2793 ./parsergen --tag calc -o calc parsergen.mdc
2794 calc : calc.o libparser.o libscanner.o libmdcode.o libnumber.o
2795 $(CC) $(CFLAGS) -o calc calc.o libparser.o libscanner.o libmdcode.o libnumber.o -licuuc -lgmp
2800 // what do we use for a demo-grammar? A calculator of course.
2812 #include <sys/mman.h>
2817 #include "scanner.h"
2823 static void free_number(struct number *n)
2829 int main(int argc, char *argv[])
2831 int fd = open(argv[1], O_RDONLY);
2832 int len = lseek(fd, 0, 2);
2833 char *file = mmap(NULL, len, PROT_READ, MAP_SHARED, fd, 0);
2834 struct section *s = code_extract(file, file+len, NULL);
2835 struct token_config config = {
2836 .ignored = (1 << TK_line_comment)
2837 | (1 << TK_block_comment)
2840 .number_chars = ".,_+-",
2844 parse_calc(s->code, &config, argc > 2 ? stderr : NULL);
2846 struct section *t = s->next;
2856 Session -> Session Line
2859 Line -> Expression NEWLINE ${ gmp_printf("Answer = %Qd\n", $1.val);
2860 { mpf_t fl; mpf_init2(fl, 20); mpf_set_q(fl, $1.val);
2861 gmp_printf(" or as a decimal: %Fg\n", fl);
2865 | Expression = Expression NEWLINE ${
2866 if (mpq_equal($1.val, $3.val))
2867 gmp_printf("Both equal %Qd\n", $1.val);
2869 gmp_printf("NOT EQUAL: %Qd\n != : %Qd\n",
2874 | NEWLINE ${ printf("Blank line\n"); }$
2875 | ERROR NEWLINE ${ printf("Skipped a bad line\n"); }$
2878 Expression -> Expression + Term ${ mpq_init($0.val); mpq_add($0.val, $1.val, $3.val); }$
2879 | Expression - Term ${ mpq_init($0.val); mpq_sub($0.val, $1.val, $3.val); }$
2880 | Term ${ mpq_init($0.val); mpq_set($0.val, $1.val); }$
2882 Term -> Term * Factor ${ mpq_init($0.val); mpq_mul($0.val, $1.val, $3.val); }$
2883 | Term / Factor ${ mpq_init($0.val); mpq_div($0.val, $1.val, $3.val); }$
2884 | Factor ${ mpq_init($0.val); mpq_set($0.val, $1.val); }$
2886 Factor -> NUMBER ${ if (number_parse($0.val, $0.tail, $1.txt) == 0) mpq_init($0.val); }$
2887 | ( Expression ) ${ mpq_init($0.val); mpq_set($0.val, $2.val); }$