// TODO(haberman): discard upb_handlers* objects that do not actually have any
// handlers set and cannot reach any upb_handlers* object that does. This is
// slightly tricky to do correctly.
static upb_handlers *newformsg(const upb_msgdef *m, const void *owner,
dfs_state *s) {
upb_handlers *h = upb_handlers_new(m, owner);
if (!h) return NULL;
if (!upb_inttable_insertptr(&s->tab, m, upb_value_ptr(h))) goto oom;
s->callback(s->closure, h);
// For each submessage field, get or create a handlers object and set it as
// the subhandlers.
upb_msg_iter i;
for(upb_msg_begin(&i, m); !upb_msg_done(&i); upb_msg_next(&i)) {
upb_fielddef *f = upb_msg_iter_field(&i);
if (!upb_fielddef_issubmsg(f)) continue;
const upb_msgdef *subdef = upb_downcast_msgdef(upb_fielddef_subdef(f));
upb_value subm_ent;
if (upb_inttable_lookupptr(&s->tab, subdef, &subm_ent)) {
upb_handlers_setsubhandlers(h, f, upb_value_getptr(subm_ent));
} else {
upb_handlers *sub_mh = newformsg(subdef, &sub_mh, s);
if (!sub_mh) goto oom;
upb_handlers_setsubhandlers(h, f, sub_mh);
upb_handlers_unref(sub_mh, &sub_mh);
}
}
return h;
oom:
upb_handlers_unref(h, owner);
return NULL;
}
bool upb_symtab_dfs(upb_def *def, upb_def **open_defs, int n,
upb_strtable *addtab) {
// This linear search makes the DFS O(n^2) in the length of the paths.
// Could make this O(n) with a hash table, but n is small.
for (int i = 0; i < n; i++) {
if (def == open_defs[i]) return false;
}
bool needcopy = false;
upb_msgdef *m = upb_dyncast_msgdef(def);
if (m) {
upb_msg_iter i;
open_defs[n++] = def;
for(i = upb_msg_begin(m); !upb_msg_done(i); i = upb_msg_next(m, i)) {
upb_fielddef *f = upb_msg_iter_field(i);
if (!upb_hassubdef(f)) continue;
needcopy |= upb_symtab_dfs(f->def, open_defs, n, addtab);
}
}
bool replacing = (upb_strtable_lookup(addtab, m->base.fqname) != NULL);
if (needcopy && !replacing) {
upb_symtab_ent e = {upb_def_dup(def)};
upb_strtable_insert(addtab, def->fqname, &e);
replacing = true;
}
return replacing;
}
static void upb_msgdef_free(upb_msgdef *m) {
upb_msg_iter i;
for(i = upb_msg_begin(m); !upb_msg_done(i); i = upb_msg_next(m, i))
upb_fielddef_free(upb_msg_iter_field(i));
upb_strtable_free(&m->ntof);
upb_inttable_free(&m->itof);
upb_def_uninit(&m->base);
free(m);
}
static void visitmsg(const upb_refcounted *r, upb_refcounted_visit *visit,
void *closure) {
const upb_msgdef *m = (const upb_msgdef*)r;
upb_msg_iter i;
for(upb_msg_begin(&i, m); !upb_msg_done(&i); upb_msg_next(&i)) {
upb_fielddef *f = upb_msg_iter_field(&i);
visit(r, upb_upcast2(f), closure);
}
}
static void validate_msgdef(const upb_msgdef* msgdef) {
// Verify that no required fields exist. proto3 does not support these.
upb_msg_iter it;
for (upb_msg_begin(&it, msgdef); !upb_msg_done(&it); upb_msg_next(&it)) {
const upb_fielddef* field = upb_msg_iter_field(&it);
if (upb_fielddef_label(field) == UPB_LABEL_REQUIRED) {
rb_raise(rb_eTypeError, "Required fields are unsupported in proto3.");
}
}
}
static void visithandlers(const upb_refcounted *r, upb_refcounted_visit *visit,
void *closure) {
const upb_handlers *h = (const upb_handlers*)r;
upb_msg_iter i;
for(upb_msg_begin(&i, h->msg); !upb_msg_done(&i); upb_msg_next(&i)) {
upb_fielddef *f = upb_msg_iter_field(&i);
if (!upb_fielddef_issubmsg(f)) continue;
const upb_handlers *sub = upb_handlers_getsubhandlers(h, f);
if (sub) visit(r, UPB_UPCAST(sub), closure);
}
}
upb_msgdef *upb_msgdef_dup(const upb_msgdef *m) {
upb_msgdef *newm = upb_msgdef_new();
newm->size = m->size;
newm->hasbit_bytes = m->hasbit_bytes;
newm->extstart = m->extstart;
newm->extend = m->extend;
upb_msg_iter i;
for(i = upb_msg_begin(m); !upb_msg_done(i); i = upb_msg_next(m, i)) {
upb_msgdef_addfield(newm, upb_fielddef_dup(upb_msg_iter_field(i)));
}
return newm;
}
/*
* call-seq:
* Descriptor.each(&block)
*
* Iterates over fields in this message type, yielding to the block on each one.
*/
VALUE Descriptor_each(VALUE _self) {
DEFINE_SELF(Descriptor, self, _self);
upb_msg_iter it;
for (upb_msg_begin(&it, self->msgdef);
!upb_msg_done(&it);
upb_msg_next(&it)) {
const upb_fielddef* field = upb_msg_iter_field(&it);
VALUE obj = get_def_obj(field);
rb_yield(obj);
}
return Qnil;
}
upb_msgdef *upb_msgdef_dup(const upb_msgdef *m, const void *owner) {
upb_msgdef *newm = upb_msgdef_new(owner);
if (!newm) return NULL;
upb_def_setfullname(upb_upcast(newm), upb_def_fullname(upb_upcast(m)));
upb_msg_iter i;
for(upb_msg_begin(&i, m); !upb_msg_done(&i); upb_msg_next(&i)) {
upb_fielddef *f = upb_fielddef_dup(upb_msg_iter_field(&i), &f);
if (!f || !upb_msgdef_addfield(newm, f, &f)) {
upb_msgdef_unref(newm, owner);
return NULL;
}
}
return newm;
}
bool upb_def_freeze(upb_def *const* defs, int n, upb_status *s) {
// First perform validation, in two passes so we can check that we have a
// transitive closure without needing to search.
for (int i = 0; i < n; i++) {
upb_def *def = defs[i];
if (upb_def_isfrozen(def)) {
// Could relax this requirement if it's annoying.
upb_status_seterrliteral(s, "def is already frozen");
goto err;
} else if (def->type == UPB_DEF_FIELD) {
upb_status_seterrliteral(s, "standalone fielddefs can not be frozen");
goto err;
} else {
// Set now to detect transitive closure in the second pass.
def->came_from_user = true;
}
}
for (int i = 0; i < n; i++) {
upb_msgdef *m = upb_dyncast_msgdef_mutable(defs[i]);
upb_enumdef *e = upb_dyncast_enumdef_mutable(defs[i]);
if (m) {
upb_inttable_compact(&m->itof);
upb_msg_iter j;
uint32_t selector = UPB_STATIC_SELECTOR_COUNT;
for(upb_msg_begin(&j, m); !upb_msg_done(&j); upb_msg_next(&j)) {
upb_fielddef *f = upb_msg_iter_field(&j);
assert(f->msgdef == m);
if (!upb_validate_field(f, s)) goto err;
f->selector_base = selector + upb_handlers_selectorbaseoffset(f);
selector += upb_handlers_selectorcount(f);
}
m->selector_count = selector;
} else if (e) {
upb_inttable_compact(&e->iton);
}
}
// Validation all passed; freeze the defs.
return upb_refcounted_freeze((upb_refcounted*const*)defs, n, s);
err:
for (int i = 0; i < n; i++) {
defs[i]->came_from_user = false;
}
assert(!upb_ok(s));
return false;
}
upb_msgdef *upb_msgdef_dup(const upb_msgdef *m, const void *owner) {
upb_msgdef *newm = upb_msgdef_new(owner);
if (!newm) return NULL;
bool ok = upb_def_setfullname(UPB_UPCAST(newm),
upb_def_fullname(UPB_UPCAST(m)), NULL);
UPB_ASSERT_VAR(ok, ok);
upb_msg_iter i;
for(upb_msg_begin(&i, m); !upb_msg_done(&i); upb_msg_next(&i)) {
upb_fielddef *f = upb_fielddef_dup(upb_msg_iter_field(&i), &f);
if (!f || !upb_msgdef_addfield(newm, f, &f, NULL)) {
upb_msgdef_unref(newm, owner);
return NULL;
}
}
return newm;
}
void upb_msgdef_layout(upb_msgdef *m) {
// Create an ordering over the fields, but only include fields with accessors.
upb_fielddef **sorted_fields =
malloc(sizeof(upb_fielddef*) * upb_msgdef_numfields(m));
int n = 0;
upb_msg_iter i;
for (i = upb_msg_begin(m); !upb_msg_done(i); i = upb_msg_next(m, i)) {
upb_fielddef *f = upb_msg_iter_field(i);
if (f->accessor) sorted_fields[n++] = f;
}
m->hasbit_bytes = upb_div_round_up(n, 8);
m->size = m->hasbit_bytes; // + header_size?
// Assign hasbits.
qsort(sorted_fields, n, sizeof(*sorted_fields), upb_fielddef_cmphasbit);
for (int i = 0; i < n; i++) {
upb_fielddef *f = sorted_fields[i];
f->hasbit = i;
}
// Assign value offsets.
qsort(sorted_fields, n, sizeof(*sorted_fields), upb_fielddef_cmpval);
size_t max_align = 0;
for (int i = 0; i < n; i++) {
upb_fielddef *f = sorted_fields[i];
const upb_type_info *type_info = &upb_types[f->type];
size_t size = type_info->size;
size_t align = type_info->align;
if (upb_isseq(f)) {
size = sizeof(void*);
align = alignof(void*);
}
// General alignment rules are: each member must be at an address that is a
// multiple of that type's alignment. Also, the size of the structure as a
// whole must be a multiple of the greatest alignment of any member.
f->offset = upb_align_up(m->size, align);
m->size = f->offset + size;
max_align = UPB_MAX(max_align, align);
}
if (max_align > 0) m->size = upb_align_up(m->size, max_align);
free(sorted_fields);
}
static upb_flow_t upb_msg_dispatch(upb_msg *msg, upb_msgdef *md,
upb_dispatcher *d) {
upb_msg_iter i;
for(i = upb_msg_begin(md); !upb_msg_done(i); i = upb_msg_next(md, i)) {
upb_fielddef *f = upb_msg_iter_field(i);
if (!upb_msg_has(msg, f)) continue;
upb_fhandlers *hf = upb_dispatcher_lookup(d, f->number);
if (!hf) continue;
upb_value val = upb_msg_get(msg, f);
if (upb_isarray(f)) {
upb_array *arr = upb_value_getarr(val);
for (uint32_t j = 0; j < upb_array_len(arr); ++j) {
upb_msg_pushval(upb_array_get(arr, f, j), f, d, hf);
}
} else {
upb_msg_pushval(val, f, d, hf);
}
}
return UPB_CONTINUE;
}
void upb_stdmsg_free(void *m, const upb_msgdef *md) {
if (m == NULL) return;
upb_msg_iter i;
for(i = upb_msg_begin(md); !upb_msg_done(i); i = upb_msg_next(md, i)) {
upb_fielddef *f = upb_msg_iter_field(i);
if (!upb_isseq(f) && !upb_issubmsg(f) && !upb_isstring(f)) continue;
void *subp = upb_value_getptr(upb_stdmsg_getptr(m, f->fval));
if (subp == NULL) continue;
if (upb_isseq(f)) {
upb_stdseq_free(subp, f);
} else if (upb_issubmsg(f)) {
upb_stdmsg_free(subp, upb_downcast_msgdef(f->def));
} else {
upb_stdarray *str = subp;
free(str->ptr);
free(str);
}
}
free(m);
}
static bool assign_msg_indices(upb_msgdef *m, upb_status *s) {
// Sort fields. upb internally relies on UPB_TYPE_MESSAGE fields having the
// lowest indexes, but we do not publicly guarantee this.
int n = upb_msgdef_numfields(m);
upb_fielddef **fields = malloc(n * sizeof(*fields));
if (!fields) return false;
upb_msg_iter j;
int i;
m->submsg_field_count = 0;
for(i = 0, upb_msg_begin(&j, m); !upb_msg_done(&j); upb_msg_next(&j), i++) {
upb_fielddef *f = upb_msg_iter_field(&j);
assert(f->msg.def == m);
if (!upb_validate_field(f, s)) {
free(fields);
return false;
}
if (upb_fielddef_issubmsg(f)) {
m->submsg_field_count++;
}
fields[i] = f;
}
qsort(fields, n, sizeof(*fields), cmp_fields);
uint32_t selector = UPB_STATIC_SELECTOR_COUNT + m->submsg_field_count;
for (i = 0; i < n; i++) {
upb_fielddef *f = fields[i];
f->index_ = i;
f->selector_base = selector + upb_handlers_selectorbaseoffset(f);
selector += upb_handlers_selectorcount(f);
}
m->selector_count = selector;
free(fields);
return false;
}
bool upb_handlers_freeze(upb_handlers *const*handlers, int n, upb_status *s) {
// TODO: verify we have a transitive closure.
for (int i = 0; i < n; i++) {
upb_handlers *h = handlers[i];
if (!upb_ok(&h->status_)) {
upb_status_seterrf(s, "handlers for message %s had error status: %s",
upb_msgdef_fullname(upb_handlers_msgdef(h)),
upb_status_errmsg(&h->status_));
return false;
}
// Check that there are no closure mismatches due to missing Start* handlers
// or subhandlers with different type-level types.
upb_msg_iter j;
for(upb_msg_begin(&j, h->msg); !upb_msg_done(&j); upb_msg_next(&j)) {
const upb_fielddef *f = upb_msg_iter_field(&j);
if (upb_fielddef_isseq(f)) {
if (!checkstart(h, f, UPB_HANDLER_STARTSEQ, s))
return false;
}
if (upb_fielddef_isstring(f)) {
if (!checkstart(h, f, UPB_HANDLER_STARTSTR, s))
return false;
}
if (upb_fielddef_issubmsg(f)) {
bool hashandler = false;
if (upb_handlers_gethandler(h, getsel(h, f, UPB_HANDLER_STARTSUBMSG)) ||
upb_handlers_gethandler(h, getsel(h, f, UPB_HANDLER_ENDSUBMSG))) {
hashandler = true;
}
if (upb_fielddef_isseq(f) &&
(upb_handlers_gethandler(h, getsel(h, f, UPB_HANDLER_STARTSEQ)) ||
upb_handlers_gethandler(h, getsel(h, f, UPB_HANDLER_ENDSEQ)))) {
hashandler = true;
}
if (hashandler && !upb_handlers_getsubhandlers(h, f)) {
// For now we add an empty subhandlers in this case. It makes the
// decoder code generator simpler, because it only has to handle two
// cases (submessage has handlers or not) as opposed to three
// (submessage has handlers in enclosing message but no subhandlers).
//
// This makes parsing less efficient in the case that we want to
// notice a submessage but skip its contents (like if we're testing
// for submessage presence or counting the number of repeated
// submessages). In this case we will end up parsing the submessage
// field by field and throwing away the results for each, instead of
// skipping the whole delimited thing at once. If this is an issue we
// can revisit it, but do remember that this only arises when you have
// handlers (startseq/startsubmsg/endsubmsg/endseq) set for the
// submessage but no subhandlers. The uses cases for this are
// limited.
upb_handlers *sub = upb_handlers_new(upb_fielddef_msgsubdef(f), &sub);
upb_handlers_setsubhandlers(h, f, sub);
upb_handlers_unref(sub, &sub);
}
// TODO(haberman): check type of submessage.
// This is slightly tricky; also consider whether we should check that
// they match at setsubhandlers time.
}
}
}
if (!upb_refcounted_freeze((upb_refcounted*const*)handlers, n, s,
UPB_MAX_HANDLER_DEPTH)) {
return false;
}
return true;
}
bool upb_symtab_add(upb_symtab *s, upb_def **defs, int n, upb_status *status) {
upb_rwlock_wrlock(&s->lock);
// Add all defs to a table for resolution.
upb_strtable addtab;
upb_strtable_init(&addtab, n, sizeof(upb_symtab_ent));
for (int i = 0; i < n; i++) {
upb_def *def = defs[i];
if (upb_strtable_lookup(&addtab, def->fqname)) {
upb_status_seterrf(status, "Conflicting defs named '%s'", def->fqname);
upb_strtable_free(&addtab);
return false;
}
upb_strtable_insert(&addtab, def->fqname, &def);
}
// All existing defs that can reach defs that are being replaced must
// themselves be replaced with versions that will point to the new defs.
// Do a DFS -- any path that finds a new def must replace all ancestors.
upb_strtable *symtab = &s->symtab;
upb_strtable_iter i;
upb_strtable_begin(&i, symtab);
for(; !upb_strtable_done(&i); upb_strtable_next(&i)) {
upb_def *open_defs[UPB_MAX_TYPE_DEPTH];
const upb_symtab_ent *e = upb_strtable_iter_value(&i);
upb_symtab_dfs(e->def, open_defs, 0, &addtab);
}
// Resolve all refs.
upb_strtable_begin(&i, &addtab);
for(; !upb_strtable_done(&i); upb_strtable_next(&i)) {
const upb_symtab_ent *e = upb_strtable_iter_value(&i);
upb_msgdef *m = upb_dyncast_msgdef(e->def);
if(!m) continue;
// Type names are resolved relative to the message in which they appear.
const char *base = m->base.fqname;
upb_msg_iter j;
for(j = upb_msg_begin(m); !upb_msg_done(j); j = upb_msg_next(m, j)) {
upb_fielddef *f = upb_msg_iter_field(j);
if (f->type == 0) {
upb_status_seterrf(status, "Field type was not set.");
return false;
}
if (!upb_hassubdef(f)) continue; // No resolving necessary.
upb_downcast_unresolveddef(f->def); // Type check.
const char *name = f->def->fqname;
// Resolve from either the addtab (pending adds) or symtab (existing
// defs). If both exist, prefer the pending add, because it will be
// overwriting the existing def.
upb_symtab_ent *found;
if(!(found = upb_resolve(&addtab, base, name)) &&
!(found = upb_resolve(symtab, base, name))) {
upb_status_seterrf(status, "could not resolve symbol '%s' "
"in context '%s'", name, base);
return false;
}
// Check the type of the found def.
upb_fieldtype_t expected = upb_issubmsg(f) ? UPB_DEF_MSG : UPB_DEF_ENUM;
if(found->def->type != expected) {
upb_status_seterrliteral(status, "Unexpected type");
return false;
}
if (!upb_fielddef_resolve(f, found->def, status)) return false;
}
}
// The defs in the transaction have been vetted, and can be moved to the
// symtab without causing errors.
upb_strtable_begin(&i, &addtab);
for(; !upb_strtable_done(&i); upb_strtable_next(&i)) {
const upb_symtab_ent *tmptab_e = upb_strtable_iter_value(&i);
upb_def_movetosymtab(tmptab_e->def, s);
upb_symtab_ent *symtab_e =
upb_strtable_lookup(&s->symtab, tmptab_e->def->fqname);
if(symtab_e) {
upb_deflist_push(&s->olddefs, symtab_e->def);
symtab_e->def = tmptab_e->def;
} else {
upb_strtable_insert(&s->symtab, tmptab_e->def->fqname, tmptab_e);
}
}
upb_strtable_free(&addtab);
upb_rwlock_unlock(&s->lock);
upb_symtab_gc(s);
return true;
}