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 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;
}
MessageLayout* create_layout(const upb_msgdef* msgdef) {
MessageLayout* layout = ALLOC(MessageLayout);
int nfields = upb_msgdef_numfields(msgdef);
upb_msg_field_iter it;
upb_msg_oneof_iter oit;
size_t off = 0;
layout->fields = ALLOC_N(MessageField, nfields);
size_t hasbit = 0;
for (upb_msg_field_begin(&it, msgdef);
!upb_msg_field_done(&it);
upb_msg_field_next(&it)) {
const upb_fielddef* field = upb_msg_iter_field(&it);
if (upb_fielddef_haspresence(field)) {
layout->fields[upb_fielddef_index(field)].hasbit = hasbit++;
} else {
layout->fields[upb_fielddef_index(field)].hasbit =
MESSAGE_FIELD_NO_HASBIT;
}
}
if (hasbit != 0) {
off += (hasbit + 8 - 1) / 8;
}
for (upb_msg_field_begin(&it, msgdef);
!upb_msg_field_done(&it);
upb_msg_field_next(&it)) {
const upb_fielddef* field = upb_msg_iter_field(&it);
size_t field_size;
if (upb_fielddef_containingoneof(field)) {
// Oneofs are handled separately below.
continue;
}
// Allocate |field_size| bytes for this field in the layout.
field_size = 0;
if (upb_fielddef_label(field) == UPB_LABEL_REPEATED) {
field_size = sizeof(VALUE);
} else {
field_size = native_slot_size(upb_fielddef_type(field));
}
// Align current offset up to |size| granularity.
off = align_up_to(off, field_size);
layout->fields[upb_fielddef_index(field)].offset = off;
layout->fields[upb_fielddef_index(field)].case_offset =
MESSAGE_FIELD_NO_CASE;
off += field_size;
}
// Handle oneofs now -- we iterate over oneofs specifically and allocate only
// one slot per oneof.
//
// We assign all value slots first, then pack the 'case' fields at the end,
// since in the common case (modern 64-bit platform) these are 8 bytes and 4
// bytes respectively and we want to avoid alignment overhead.
//
// Note that we reserve 4 bytes (a uint32) per 'case' slot because the value
// space for oneof cases is conceptually as wide as field tag numbers. In
// practice, it's unlikely that a oneof would have more than e.g. 256 or 64K
// members (8 or 16 bits respectively), so conceivably we could assign
// consecutive case numbers and then pick a smaller oneof case slot size, but
// the complexity to implement this indirection is probably not worthwhile.
for (upb_msg_oneof_begin(&oit, msgdef);
!upb_msg_oneof_done(&oit);
upb_msg_oneof_next(&oit)) {
const upb_oneofdef* oneof = upb_msg_iter_oneof(&oit);
upb_oneof_iter fit;
// Always allocate NATIVE_SLOT_MAX_SIZE bytes, but share the slot between
// all fields.
size_t field_size = NATIVE_SLOT_MAX_SIZE;
// Align the offset.
off = align_up_to(off, field_size);
// Assign all fields in the oneof this same offset.
for (upb_oneof_begin(&fit, oneof);
!upb_oneof_done(&fit);
upb_oneof_next(&fit)) {
const upb_fielddef* field = upb_oneof_iter_field(&fit);
layout->fields[upb_fielddef_index(field)].offset = off;
}
off += field_size;
}
// Now the case fields.
for (upb_msg_oneof_begin(&oit, msgdef);
!upb_msg_oneof_done(&oit);
upb_msg_oneof_next(&oit)) {
const upb_oneofdef* oneof = upb_msg_iter_oneof(&oit);
upb_oneof_iter fit;
size_t field_size = sizeof(uint32_t);
// Align the offset.
off = (off + field_size - 1) & ~(field_size - 1);
// Assign all fields in the oneof this same offset.
for (upb_oneof_begin(&fit, oneof);
!upb_oneof_done(&fit);
upb_oneof_next(&fit)) {
const upb_fielddef* field = upb_oneof_iter_field(&fit);
layout->fields[upb_fielddef_index(field)].case_offset = off;
}
off += field_size;
}
layout->size = off;
layout->msgdef = msgdef;
upb_msgdef_ref(layout->msgdef, &layout->msgdef);
return layout;
}
static bool upb_msglayout_init(const upb_msgdef *m,
upb_msglayout *l,
upb_msgfactory *factory) {
upb_msg_field_iter it;
upb_msg_oneof_iter oit;
size_t hasbit;
size_t submsg_count = 0;
const upb_msglayout **submsgs;
upb_msglayout_field *fields;
for (upb_msg_field_begin(&it, m);
!upb_msg_field_done(&it);
upb_msg_field_next(&it)) {
const upb_fielddef* f = upb_msg_iter_field(&it);
if (upb_fielddef_issubmsg(f)) {
submsg_count++;
}
}
memset(l, 0, sizeof(*l));
fields = upb_gmalloc(upb_msgdef_numfields(m) * sizeof(*fields));
submsgs = upb_gmalloc(submsg_count * sizeof(*submsgs));
if ((!fields && upb_msgdef_numfields(m)) ||
(!submsgs && submsg_count)) {
/* OOM. */
upb_gfree(fields);
upb_gfree(submsgs);
return false;
}
l->field_count = upb_msgdef_numfields(m);
l->fields = fields;
l->submsgs = submsgs;
/* Allocate data offsets in three stages:
*
* 1. hasbits.
* 2. regular fields.
* 3. oneof fields.
*
* OPT: There is a lot of room for optimization here to minimize the size.
*/
/* Allocate hasbits and set basic field attributes. */
submsg_count = 0;
for (upb_msg_field_begin(&it, m), hasbit = 0;
!upb_msg_field_done(&it);
upb_msg_field_next(&it)) {
const upb_fielddef* f = upb_msg_iter_field(&it);
upb_msglayout_field *field = &fields[upb_fielddef_index(f)];
field->number = upb_fielddef_number(f);
field->descriptortype = upb_fielddef_descriptortype(f);
field->label = upb_fielddef_label(f);
if (upb_fielddef_issubmsg(f)) {
const upb_msglayout *sub_layout =
upb_msgfactory_getlayout(factory, upb_fielddef_msgsubdef(f));
field->submsg_index = submsg_count++;
submsgs[field->submsg_index] = sub_layout;
}
if (upb_fielddef_haspresence(f) && !upb_fielddef_containingoneof(f)) {
field->presence = (hasbit++);
} else {
field->presence = 0;
}
}
/* Account for space used by hasbits. */
l->size = div_round_up(hasbit, 8);
/* Allocate non-oneof fields. */
for (upb_msg_field_begin(&it, m); !upb_msg_field_done(&it);
upb_msg_field_next(&it)) {
const upb_fielddef* f = upb_msg_iter_field(&it);
size_t field_size = upb_msg_fielddefsize(f);
size_t index = upb_fielddef_index(f);
if (upb_fielddef_containingoneof(f)) {
/* Oneofs are handled separately below. */
continue;
}
fields[index].offset = upb_msglayout_place(l, field_size);
}
/* Allocate oneof fields. Each oneof field consists of a uint32 for the case
* and space for the actual data. */
for (upb_msg_oneof_begin(&oit, m); !upb_msg_oneof_done(&oit);
upb_msg_oneof_next(&oit)) {
const upb_oneofdef* o = upb_msg_iter_oneof(&oit);
upb_oneof_iter fit;
size_t case_size = sizeof(uint32_t); /* Could potentially optimize this. */
size_t field_size = 0;
uint32_t case_offset;
uint32_t data_offset;
/* Calculate field size: the max of all field sizes. */
for (upb_oneof_begin(&fit, o);
!upb_oneof_done(&fit);
upb_oneof_next(&fit)) {
const upb_fielddef* f = upb_oneof_iter_field(&fit);
field_size = UPB_MAX(field_size, upb_msg_fielddefsize(f));
}
/* Align and allocate case offset. */
case_offset = upb_msglayout_place(l, case_size);
data_offset = upb_msglayout_place(l, field_size);
for (upb_oneof_begin(&fit, o);
!upb_oneof_done(&fit);
upb_oneof_next(&fit)) {
const upb_fielddef* f = upb_oneof_iter_field(&fit);
fields[upb_fielddef_index(f)].offset = data_offset;
fields[upb_fielddef_index(f)].presence = ~case_offset;
}
}
/* Size of the entire structure should be a multiple of its greatest
* alignment. TODO: track overall alignment for real? */
l->size = align_up(l->size, 8);
return true;
}