+###### heart of parser
+
+ shift(&p, TK_eof, NULL, states);
+ while (!accepted && p.tos > 0) {
+ struct frame *tos = &p.stack[p.tos-1];
+ if (!tk)
+ tk = tok_copy(token_next(tokens));
+ parser_trace(trace, &p,
+ tk, states, non_term, config->known_count);
+
+ ## try shift or ignore
+ ## try reduce
+ ## handle error
+ }
+
+Indents are ignored unless they can be shifted onto the stack
+immediately or nothing can be shifted (in which case we reduce, and try
+again). The end of an indented section - the OUT token - is ignored
+precisely when the indent was ignored. To keep track of this we need a
+small stack of flags, which is easily stored as bits in an `unsigned
+long`. This will never overflow and the scanner only allows 20 levels
+of indentation.
+
+###### parser state
+ unsigned long ignored_indents;
+
+NEWLINE is ignored when in an indented section of text which was not
+explicitly expected by the grammar. So if the most recent indent is
+ignored, so is any NEWLINE token.
+
+If a NEWLINE is seen but it cannot be shifted, we try to shift an EOL
+token instead. If that succeeds, we make a new copy of the NEWLINE
+token and continue. This allows a NEWLINE to appear to be preceded by
+an indefinite number of EOL tokens.
+
+The token number for `EOL` cannot be statically declared, so when the
+parser starts we need to look through the array of non-terminals to find
+the EOL.
+
+###### parser state
+ int tk_eol;
+
+###### find eol
+ p.tk_eol = 0;
+ while (strcmp(non_term[p.tk_eol], "EOL") != 0)
+ p.tk_eol += 1;
+ p.tk_eol += TK_reserved + config->known_count;
+
+For other tokens, we shift the next token if that is possible, otherwise
+we try to reduce a production.
+
+###### try shift or ignore
+
+ if ((tk->num == TK_newline || tk->num == TK_out) &&
+ (p.ignored_indents & 1)) {
+ /* indented, so ignore OUT and NEWLINE */
+ if (tk->num == TK_out)
+ p.ignored_indents >>= 1;
+ free(tk);
+ tk = NULL;
+ parser_trace_action(trace, "Ignore");
+ continue;
+ }
+
+ if (shift(&p, tk->num, tk, states)) {
+ if (tk->num == TK_out)
+ p.ignored_indents >>= 1;
+ if (tk->num == TK_in)
+ p.ignored_indents <<= 1;
+
+ parser_trace_action(trace, "Shift");
+ tk = NULL;
+ ## did shift
+ continue;
+ } else if (tk->num == TK_newline &&
+ shift(&p, p.tk_eol, tk, states)) {
+ tk = tok_copy(*tk);
+ parser_trace_action(trace, "ShiftEOL");
+ continue;
+ }
+
+ if (tk->num == TK_in && states[p.stack[p.tos-1].state].go_to_cnt > 0) {
+ /* No indent expected here and reduce is not mandatory, so ignore IN */
+ free(tk);
+ tk = NULL;
+ p.ignored_indents <<= 1;
+ p.ignored_indents |= 1;
+ parser_trace_action(trace, "Ignore");
+ continue;
+ }
+
+We have already discussed the bulk of the handling of a "reduce" action,
+with the `pop()` and `shift()` functions doing much of the work. There
+is a little more complexity needed to manage storage for the asn (Abstract
+Syntax Node), and also a test of whether the reduction is permitted.
+
+When we try to shift the result of reducing production-zero, it will
+fail because there is no next state. In this case the asn will not have
+been stored on the stack, so it get stored in the `ret` variable, and we
+report that that input has been accepted.
+
+###### parser vars
+
+ void *ret = NULL;
+
+###### try reduce
+
+ if (states[tos->state].reduce_prod >= 0) {
+ void **body;
+ void *res;
+ const struct state *nextstate = &states[tos->state];
+ int prod = nextstate->reduce_prod;
+ int size = reductions[prod].size;
+ int res_size = reductions[prod].result_size;
+
+ body = p.asn_stack + (p.tos - size);
+ res = res_size ? calloc(1, res_size) : NULL;
+ res_size = do_reduce(prod, body, config, res);
+ if (res_size != reductions[prod].result_size)
+ abort();
+ pop(&p, size, do_free);
+ if (!shift(&p, reductions[prod].sym, res, states)) {
+ accepted = 1;
+ ret = res;
+ parser_trace_action(trace, "Accept");
+ } else
+ parser_trace_action(trace, "Reduce");
+ continue;
+ }
+
+If we can neither shift nor reduce we have an error to handle. There
+are two possible responses to an error: we can pop single frames off the
+stack until we find a frame that can shift `TK_error`, or we can discard
+the current look-ahead symbol.
+
+When we first see an error we do the first of these and set a flag to
+record that we are processing an error. If the normal shift/reduce
+tests fail to find that anything can be done from that state, we start
+dropping tokens until either we manage to shift one, or reach end-of-file.
+
+###### parser vars
+
+ int in_err = 0;
+
+###### did shift
+
+ in_err = 0;
+
+###### handle error
+
+ if (in_err) {
+ parser_trace_action(trace, "DISCARD");
+ if (tk->num == TK_eof)
+ break;
+ free(tk);
+ tk = NULL;
+ } else {
+ struct token *err_tk;
+
+ parser_trace_action(trace, "ERROR");
+
+ err_tk = tok_copy(*tk);
+ while (p.tos > 0 && shift(&p, TK_error, err_tk, states) == 0)
+ // discard this state
+ pop(&p, 1, do_free);
+ if (p.tos == 0) {
+ free(err_tk);
+ // no state accepted TK_error
+ break;
code / ocean / blobdiff