Tables are compressed, but perf goes down to 2.44GB/s.

BUILD

@@ -42,6 +42,7 @@ COPTS = CPPOPTS + [
     # copybara:strip_for_google3_begin
     #"-pedantic",
     #"-Werror=pedantic",
+    "-std=gnu11",
     "-Wstrict-prototypes",
     # copybara:strip_end
 ]

upb/decode_fast.c

@@ -28,9 +28,11 @@ const char *fastdecode_tag_dispatch(upb_decstate *d, const char *ptr, upb_msg *m
                                 const upb_msglayout *table, uint64_t hasbits, uint32_t tag) {
   uint64_t data;
   size_t idx;
-  idx = (tag & 0xf8) >> 3;
-  data = table->field_data[idx] ^ tag;
-  return table->field_parser[idx](UPB_PARSE_ARGS);
+  idx = (tag & table->table_mask);
+  __builtin_assume((idx & 7) == 0);
+  idx >>= 3;
+  data = table->fasttable[idx].field_data ^ tag;
+  return table->fasttable[idx].field_parser(UPB_PARSE_ARGS);
 }
 
 UPB_FORCEINLINE
@@ -40,7 +42,7 @@ uint32_t fastdecode_load_tag(const char* ptr) {
   return tag;
 }
 
-UPB_FORCEINLINE
+UPB_NOINLINE
 const char *fastdecode_dispatch(upb_decstate *d, const char *ptr, upb_msg *msg,
                                 const upb_msglayout *table, uint64_t hasbits) {
   if (UPB_UNLIKELY(ptr >= d->fastlimit)) {

upb/def.c

@@ -940,9 +940,8 @@ static bool make_layout(const upb_symtab *symtab, const upb_msgdef *m) {
   const upb_msglayout **submsgs;
   upb_msglayout_field *fields;
   upb_alloc *alloc = upb_arena_alloc(symtab->arena);
-  size_t i;
 
-  memset(l, 0, sizeof(*l));
+  memset(l, 0, sizeof(*l) + sizeof(_upb_fasttable_entry));
 
   fields = upb_malloc(alloc, upb_msgdef_numfields(m) * sizeof(*fields));
   submsgs = upb_malloc(alloc, submsg_count * sizeof(*submsgs));
@@ -956,10 +955,10 @@ static bool make_layout(const upb_symtab *symtab, const upb_msgdef *m) {
   l->field_count = upb_msgdef_numfields(m);
   l->fields = fields;
   l->submsgs = submsgs;
+  l->table_mask = 0;
 
-  for (i = 0; i < 32; i++) {
-    l->field_parser[i] = &fastdecode_generic;
-  }
+  l->fasttable[0].field_parser = &fastdecode_generic;
+  l->fasttable[0].field_data = 0;
 
   if (upb_msgdef_mapentry(m)) {
     /* TODO(haberman): refactor this method so this special case is more
@@ -1722,7 +1721,8 @@ static bool create_msgdef(symtab_addctx *ctx, const char *prefix,
     ctx->layouts++;
   } else {
     /* Allocate now (to allow cross-linking), populate later. */
-    m->layout = upb_malloc(ctx->alloc, sizeof(*m->layout));
+    m->layout = upb_malloc(ctx->alloc,
+                           sizeof(*m->layout) + sizeof(_upb_fasttable_entry));
   }
 
   oneofs = google_protobuf_DescriptorProto_oneof_decl(msg_proto, &n);

upb/msg.h

@@ -54,16 +54,21 @@ typedef const char *_upb_field_parser(struct upb_decstate *d, const char *ptr,
                                       const struct upb_msglayout *table,
                                       uint64_t hasbits, uint64_t data);
 
+typedef struct {
+  _upb_field_parser *field_parser;
+  uint64_t field_data;
+} _upb_fasttable_entry;
+
 typedef struct upb_msglayout {
-  _upb_field_parser *field_parser[32];
-  uint64_t field_data[32];
   const struct upb_msglayout *const* submsgs;
   const upb_msglayout_field *fields;
+  uint32_t table_mask;
   /* Must be aligned to sizeof(void*).  Doesn't include internal members like
    * unknown fields, extension dict, pointer to msglayout, etc. */
   uint16_t size;
   uint16_t field_count;
   bool extendable;
+  _upb_fasttable_entry fasttable[];
 } upb_msglayout;
 
 /** upb_msg *******************************************************************/

upbc/generator.cc

@@ -857,11 +857,19 @@ void TryFillTableEntry(const protobuf::Descriptor* message,
 
 std::vector<TableEntry> FastDecodeTable(const protobuf::Descriptor* message,
                                         const MessageLayout& layout) {
+  int table_size = 1;
+  for (int i = 0; i < message->field_count(); i++) {
+    int num = message->field(i)->number();
+    while (num >= table_size && num < 32) {
+      table_size *= 2;
+    }
+  }
+
   std::vector<TableEntry> table;
   MessageLayout::Size empty_size;
   empty_size.size32 = 0;
   empty_size.size64 = 0;
-  for (int i = 0; i < 32; i++) {
+  for (int i = 0; i < table_size; i++) {
     table.emplace_back(TableEntry{"fastdecode_generic", empty_size});
     TryFillTableEntry(message, layout, i, table.back());
   }
@@ -963,22 +971,18 @@ void WriteSource(const protobuf::FileDescriptor* file, Output& output) {
     std::vector<TableEntry> table = FastDecodeTable(message, layout);
 
     output("const upb_msglayout $0 = {\n", MessageInit(message));
-    output("  {\n");
-    for (const auto& ent : table) {
-      output("    &$0,\n", ent.first);
-    }
-    output("  },\n");
-    output("  {\n");
-    for (const auto& ent : table) {
-      output("    $0,\n", GetSizeInit(ent.second));
-    }
-    output("  },\n");
     output("  $0,\n", submsgs_array_ref);
     output("  $0,\n", fields_array_ref);
+    output("  $0,\n", (table.size() - 1) << 3);
     output("  $0, $1, $2,\n", GetSizeInit(layout.message_size()),
            field_number_order.size(),
            "false"  // TODO: extendable
     );
+    output("  {\n");
+    for (const auto& ent : table) {
+      output("    {&$0, $1},\n", ent.first, GetSizeInit(ent.second));
+    }
+    output("  },\n");
 
     output("};\n\n");
   }