grpc 1.72.0 → 1.74.0.pre2

data/third_party/upb/upb/hash/common.h

@@ -27,10 +27,10 @@
 #ifndef UPB_HASH_COMMON_H_
 #define UPB_HASH_COMMON_H_
 
+#include <stdint.h>
 #include <string.h>
 
 #include "upb/base/string_view.h"
-#include "upb/mem/arena.h"
 
 // Must be last.
 #include "upb/port/def.inc"
@@ -45,8 +45,6 @@ typedef struct {
   uint64_t val;
 } upb_value;
 
-UPB_INLINE void _upb_value_setval(upb_value* v, uint64_t val) { v->val = val; }
-
 /* For each value ctype, define the following set of functions:
  *
  * // Get/set an int32 from a upb_value.
@@ -57,7 +55,7 @@ UPB_INLINE void _upb_value_setval(upb_value* v, uint64_t val) { v->val = val; }
  * upb_value upb_value_int32(int32_t val); */
 #define FUNCS(name, membername, type_t, converter)                   \
   UPB_INLINE void upb_value_set##name(upb_value* val, type_t cval) { \
-    val->val = (converter)cval;                                      \
+    val->val = (uint64_t)cval;                                       \
   }                                                                  \
   UPB_INLINE upb_value upb_value_##name(type_t val) {                \
     upb_value ret;                                                   \
@@ -100,45 +98,33 @@ UPB_INLINE upb_value upb_value_double(double cval) {
   return ret;
 }
 
-/* upb_tabkey *****************************************************************/
+/* upb_key *****************************************************************/
+
+// A uint32 size followed by that number of bytes stored contiguously.
+typedef struct {
+  uint32_t size;
+  const char data[];
+} upb_SizePrefixString;
 
 /* Either:
- *   1. an actual integer key, or
- *   2. a pointer to a string prefixed by its uint32_t length, owned by us.
+ *   1. an actual integer key
+ *   2. a SizePrefixString*, owned by us.
  *
- * ...depending on whether this is a string table or an int table.  We would
- * make this a union of those two types, but C89 doesn't support statically
- * initializing a non-first union member. */
-typedef uintptr_t upb_tabkey;
-
-UPB_INLINE char* upb_tabstr(upb_tabkey key, uint32_t* len) {
-  char* mem = (char*)key;
-  if (len) memcpy(len, mem, sizeof(*len));
-  return mem + sizeof(*len);
-}
-
-UPB_INLINE upb_StringView upb_tabstrview(upb_tabkey key) {
-  upb_StringView ret;
-  uint32_t len;
-  ret.data = upb_tabstr(key, &len);
-  ret.size = len;
-  return ret;
+ * ...depending on whether this is a string table or an int table. */
+typedef union {
+  uintptr_t num;
+  const upb_SizePrefixString* str;
+} upb_key;
+
+UPB_INLINE upb_StringView upb_key_strview(upb_key key) {
+  return upb_StringView_FromDataAndSize(key.str->data, key.str->size);
 }
 
-/* upb_tabval *****************************************************************/
-
-typedef struct upb_tabval {
-  uint64_t val;
-} upb_tabval;
-
-#define UPB_TABVALUE_EMPTY_INIT \
-  { -1 }
-
 /* upb_table ******************************************************************/
 
 typedef struct _upb_tabent {
-  upb_tabkey key;
-  upb_tabval val;
+  upb_value val;
+  upb_key key;
 
   /* Internal chaining.  This is const so we can create static initializers for
    * tables.  We cast away const sometimes, but *only* when the containing
@@ -148,20 +134,35 @@ typedef struct _upb_tabent {
 } upb_tabent;
 
 typedef struct {
-  size_t count;       /* Number of entries in the hash part. */
-  uint32_t mask;      /* Mask to turn hash value -> bucket. */
-  uint32_t max_count; /* Max count before we hit our load limit. */
-  uint8_t size_lg2;   /* Size of the hashtable part is 2^size_lg2 entries. */
   upb_tabent* entries;
+  /* Number of entries in the hash part. */
+  uint32_t count;
+
+  /* Mask to turn hash value -> bucket. The map's allocated size is mask + 1.*/
+  uint32_t mask;
 } upb_table;
 
-UPB_INLINE size_t upb_table_size(const upb_table* t) {
-  return t->size_lg2 ? 1 << t->size_lg2 : 0;
-}
+UPB_INLINE size_t upb_table_size(const upb_table* t) { return t->mask + 1; }
 
 // Internal-only functions, in .h file only out of necessity.
 
-UPB_INLINE bool upb_tabent_isempty(const upb_tabent* e) { return e->key == 0; }
+UPB_INLINE upb_key upb_key_empty(void) {
+  upb_key ret;
+  memset(&ret, 0, sizeof(upb_key));
+  return ret;
+}
+
+UPB_INLINE bool upb_tabent_isempty(const upb_tabent* e) {
+  upb_key key = e->key;
+  UPB_ASSERT(sizeof(key.num) == sizeof(key.str));
+  uintptr_t val;
+  memcpy(&val, &key, sizeof(val));
+  // Note: for upb_inttables a tab_key is a true integer key value, but the
+  // inttable maintains the invariant that 0 value is always stored in the
+  // compact table and never as a upb_tabent* so we can always use the 0
+  // key value to identify an empty tabent.
+  return val == 0;
+}
 
 uint32_t _upb_Hash(const void* p, size_t n, uint64_t seed);
 

data/third_party/upb/upb/hash/int_table.h

@@ -8,16 +8,28 @@
 #ifndef UPB_HASH_INT_TABLE_H_
 #define UPB_HASH_INT_TABLE_H_
 
+#include <stddef.h>
+#include <stdint.h>
+
 #include "upb/hash/common.h"
+#include "upb/mem/arena.h"
 
 // Must be last.
 #include "upb/port/def.inc"
 
 typedef struct {
-  upb_table t;              // For entries that don't fit in the array part.
-  const upb_tabval* array;  // Array part of the table. See const note above.
-  size_t array_size;        // Array part size.
-  size_t array_count;       // Array part number of elements.
+  upb_table t;  // For entries that don't fit in the array part.
+  // Array part of the table.
+  // Pointers on this table are const so we can create static initializers for
+  // tables.  We cast away const sometimes, but *only* when the containing
+  // upb_table is known to be non-const.  This requires a bit of care, but
+  // the subtlety is confined to table.c.
+  const upb_value* array;
+  // Track presence in the array part. Each bit at index (key % 8) at the
+  // presence_mask[key/8] indicates if the element is present in the array part.
+  const uint8_t* presence_mask;
+  uint32_t array_size;   // Array part size.
+  uint32_t array_count;  // Array part number of elements.
 } upb_inttable;
 
 #ifdef __cplusplus
@@ -56,7 +68,11 @@ bool upb_inttable_replace(upb_inttable* t, uintptr_t key, upb_value val);
 // Optimizes the table for the current set of entries, for both memory use and
 // lookup time. Client should call this after all entries have been inserted;
 // inserting more entries is legal, but will likely require a table resize.
-void upb_inttable_compact(upb_inttable* t, upb_Arena* a);
+// Returns false if reallocation fails.
+bool upb_inttable_compact(upb_inttable* t, upb_Arena* a);
+
+// Clears the table.
+void upb_inttable_clear(upb_inttable* t);
 
 // Iteration over inttable:
 //
@@ -72,6 +88,14 @@ void upb_inttable_compact(upb_inttable* t, upb_Arena* a);
 bool upb_inttable_next(const upb_inttable* t, uintptr_t* key, upb_value* val,
                        intptr_t* iter);
 void upb_inttable_removeiter(upb_inttable* t, intptr_t* iter);
+void upb_inttable_setentryvalue(upb_inttable* t, intptr_t iter, upb_value v);
+bool upb_inttable_done(const upb_inttable* t, intptr_t i);
+uintptr_t upb_inttable_iter_key(const upb_inttable* t, intptr_t iter);
+upb_value upb_inttable_iter_value(const upb_inttable* t, intptr_t iter);
+
+UPB_INLINE bool upb_inttable_arrhas(const upb_inttable* t, uintptr_t key) {
+  return (t->presence_mask[key / 8] & (1 << (key % 8))) != 0;
+}
 
 #ifdef __cplusplus
 } /* extern "C" */

data/third_party/upb/upb/hash/str_table.h

@@ -8,7 +8,13 @@
 #ifndef UPB_HASH_STR_TABLE_H_
 #define UPB_HASH_STR_TABLE_H_
 
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+#include "upb/base/string_view.h"
 #include "upb/hash/common.h"
+#include "upb/mem/arena.h"
 
 // Must be last.
 #include "upb/port/def.inc"

data/third_party/upb/upb/mem/arena.c

@@ -29,7 +29,9 @@ void upb_Arena_SetMaxBlockSize(size_t max) {
 
 typedef struct upb_MemBlock {
   struct upb_MemBlock* next;
-  size_t size;
+  // If this block is the head of the list, tracks a growing hint of what the
+  // *next* block should be; otherwise tracks the size of the actual allocation.
+  size_t size_or_hint;
   // Data follows.
 } upb_MemBlock;
 
@@ -55,10 +57,13 @@ typedef struct upb_ArenaInternal {
   // == NULL at end of list.
   UPB_ATOMIC(struct upb_ArenaInternal*) next;
 
-  // If the low bit is set, is a pointer to the tail of the list (populated for
-  // roots, set to self for roots with no fused arenas). If the low bit is not
-  // set, is a pointer to the previous node in the list, such that
-  // a->previous_or_tail->next == a.
+  // - If the low bit is set, is a pointer to the tail of the list (populated
+  //   for roots, set to self for roots with no fused arenas). This is best
+  //   effort, and it may not always reflect the true tail, but it will always
+  //   be a valid node in the list. This is useful for finding the list tail
+  //   without having to walk the entire list.
+  // - If the low bit is not set, is a pointer to the previous node in the list,
+  //   such that a->previous_or_tail->next == a.
   UPB_ATOMIC(uintptr_t) previous_or_tail;
 
   // Linked list of blocks to free/cleanup.
@@ -89,6 +94,12 @@ static upb_ArenaInternal* upb_Arena_Internal(const upb_Arena* a) {
   return &((upb_ArenaState*)a)->body;
 }
 
+// Extracts the (upb_Arena*) from a (upb_ArenaInternal*)
+static upb_Arena* upb_Arena_FromInternal(const upb_ArenaInternal* ai) {
+  ptrdiff_t offset = -offsetof(upb_ArenaState, body);
+  return UPB_PTR_AT(ai, offset, upb_Arena);
+}
+
 static bool _upb_Arena_IsTaggedRefcount(uintptr_t parent_or_count) {
   return (parent_or_count & 1) == 1;
 }
@@ -283,31 +294,39 @@ uint32_t upb_Arena_DebugRefCount(const upb_Arena* a) {
   return (uint32_t)_upb_Arena_RefCountFromTagged(tagged);
 }
 
+// Adds an allocated block to the head of the list.
 static void _upb_Arena_AddBlock(upb_Arena* a, void* ptr, size_t offset,
                                 size_t block_size) {
   upb_ArenaInternal* ai = upb_Arena_Internal(a);
   upb_MemBlock* block = ptr;
 
-  block->size = block_size;
-  // Insert into linked list.
-  block->next = ai->blocks;
-  ai->blocks = block;
-
+  block->size_or_hint = block_size;
   UPB_ASSERT(offset >= kUpb_MemblockReserve);
-  a->UPB_PRIVATE(ptr) = UPB_PTR_AT(block, offset, char);
+  char* start = UPB_PTR_AT(block, offset, char);
+  upb_MemBlock* head = ai->blocks;
+  if (head && head->next) {
+    // Fix up size to match actual allocation size
+    head->size_or_hint = a->UPB_PRIVATE(end) - (char*)head;
+  }
+  block->next = head;
+  ai->blocks = block;
+  a->UPB_PRIVATE(ptr) = start;
   a->UPB_PRIVATE(end) = UPB_PTR_AT(block, block_size, char);
-
-  UPB_POISON_MEMORY_REGION(a->UPB_PRIVATE(ptr),
-                           a->UPB_PRIVATE(end) - a->UPB_PRIVATE(ptr));
+  UPB_POISON_MEMORY_REGION(start, a->UPB_PRIVATE(end) - start);
+  UPB_ASSERT(UPB_PRIVATE(_upb_ArenaHas)(a) >= block_size - offset);
 }
 
-static bool _upb_Arena_AllocBlock(upb_Arena* a, size_t size) {
+// Fulfills the allocation request by allocating a new block. Returns NULL on
+// allocation failure.
+void* UPB_PRIVATE(_upb_Arena_SlowMalloc)(upb_Arena* a, size_t size) {
   upb_ArenaInternal* ai = upb_Arena_Internal(a);
-  if (!ai->block_alloc) return false;
+  if (!ai->block_alloc) return NULL;
   size_t last_size = 128;
+  size_t current_free = 0;
   upb_MemBlock* last_block = ai->blocks;
   if (last_block) {
     last_size = a->UPB_PRIVATE(end) - (char*)last_block;
+    current_free = a->UPB_PRIVATE(end) - a->UPB_PRIVATE(ptr);
   }
 
   // Relaxed order is safe here as we don't need any ordering with the setter.
@@ -315,17 +334,45 @@ static bool _upb_Arena_AllocBlock(upb_Arena* a, size_t size) {
       upb_Atomic_Load(&g_max_block_size, memory_order_relaxed);
 
   // Don't naturally grow beyond the max block size.
-  size_t clamped_size = UPB_MIN(last_size * 2, max_block_size);
-
+  size_t target_size = UPB_MIN(last_size * 2, max_block_size);
+  size_t future_free = UPB_MAX(size, target_size - kUpb_MemblockReserve) - size;
+  // We want to preserve exponential growth in block size without wasting too
+  // much unused space at the end of blocks. Once the head of our blocks list is
+  // large enough to always trigger a max-sized block for all subsequent
+  // allocations, allocate blocks that would net reduce free space behind it.
+  if (last_block && current_free > future_free &&
+      target_size < max_block_size) {
+    last_size = last_block->size_or_hint;
+    // Recalculate sizes with possibly larger last_size
+    target_size = UPB_MIN(last_size * 2, max_block_size);
+    future_free = UPB_MAX(size, target_size - kUpb_MemblockReserve) - size;
+  }
+  bool insert_after_head = false;
+  // Only insert after head if an allocated block is present; we don't want to
+  // continue allocating out of the initial block because we'll have no way of
+  // restoring the size of our allocated block if we add another.
+  if (last_block && current_free >= future_free) {
+    // If we're still going to net reduce free space with this new block, then
+    // only allocate the precise size requested and keep the current last block
+    // as the active block for future allocations.
+    insert_after_head = true;
+    target_size = size + kUpb_MemblockReserve;
+    // Add something to our previous size each time, so that eventually we
+    // will reach the max block size. Allocations larger than the max block size
+    // will always get their own backing allocation, so don't include them.
+    if (target_size <= max_block_size) {
+      last_block->size_or_hint =
+          UPB_MIN(last_block->size_or_hint + (size >> 1), max_block_size >> 1);
+    }
+  }
   // We may need to exceed the max block size if the user requested a large
   // allocation.
-  size_t block_size = UPB_MAX(kUpb_MemblockReserve + size, clamped_size);
+  size_t block_size = UPB_MAX(kUpb_MemblockReserve + size, target_size);
 
   upb_MemBlock* block =
       upb_malloc(_upb_ArenaInternal_BlockAlloc(ai), block_size);
 
-  if (!block) return false;
-  _upb_Arena_AddBlock(a, block, kUpb_MemblockReserve, block_size);
+  if (!block) return NULL;
   // Atomic add not required here, as threads won't race allocating blocks, plus
   // atomic fetch-add is slower than load/add/store on arm devices compiled
   // targetting pre-v8.1. Relaxed order is safe as nothing depends on order of
@@ -335,13 +382,21 @@ static bool _upb_Arena_AllocBlock(upb_Arena* a, size_t size) {
       upb_Atomic_Load(&ai->space_allocated, memory_order_relaxed);
   upb_Atomic_Store(&ai->space_allocated, old_space_allocated + block_size,
                    memory_order_relaxed);
-  UPB_ASSERT(UPB_PRIVATE(_upb_ArenaHas)(a) >= size);
-  return true;
-}
-
-void* UPB_PRIVATE(_upb_Arena_SlowMalloc)(upb_Arena* a, size_t size) {
-  if (!_upb_Arena_AllocBlock(a, size)) return NULL;  // OOM
-  return upb_Arena_Malloc(a, size - UPB_ASAN_GUARD_SIZE);
+  if (UPB_UNLIKELY(insert_after_head)) {
+    upb_ArenaInternal* ai = upb_Arena_Internal(a);
+    block->size_or_hint = block_size;
+    upb_MemBlock* head = ai->blocks;
+    block->next = head->next;
+    head->next = block;
+
+    char* allocated = UPB_PTR_AT(block, kUpb_MemblockReserve, char);
+    UPB_POISON_MEMORY_REGION(allocated + size, UPB_ASAN_GUARD_SIZE);
+    return allocated;
+  } else {
+    _upb_Arena_AddBlock(a, block, kUpb_MemblockReserve, block_size);
+    UPB_ASSERT(UPB_PRIVATE(_upb_ArenaHas)(a) >= size);
+    return upb_Arena_Malloc(a, size - UPB_ASAN_GUARD_SIZE);
+  }
 }
 
 static upb_Arena* _upb_Arena_InitSlow(upb_alloc* alloc, size_t first_size) {
@@ -420,18 +475,22 @@ static void _upb_Arena_DoFree(upb_ArenaInternal* ai) {
     // Load first since arena itself is likely from one of its blocks.
     upb_ArenaInternal* next_arena =
         (upb_ArenaInternal*)upb_Atomic_Load(&ai->next, memory_order_acquire);
-    // Freeing may have memory barriers that confuse tsan, so assert immdiately
+    // Freeing may have memory barriers that confuse tsan, so assert immediately
     // after load here
     if (next_arena) {
       UPB_TSAN_CHECK_PUBLISHED(next_arena);
     }
     upb_alloc* block_alloc = _upb_ArenaInternal_BlockAlloc(ai);
     upb_MemBlock* block = ai->blocks;
+    if (block && block->next) {
+      block->size_or_hint =
+          upb_Arena_FromInternal(ai)->UPB_PRIVATE(end) - (char*)block;
+    }
     upb_AllocCleanupFunc* alloc_cleanup = *ai->upb_alloc_cleanup;
     while (block != NULL) {
       // Load first since we are deleting block.
       upb_MemBlock* next_block = block->next;
-      upb_free_sized(block_alloc, block, block->size);
+      upb_free_sized(block_alloc, block, block->size_or_hint);
       block = next_block;
     }
     if (alloc_cleanup != NULL) {
@@ -477,48 +536,77 @@ retry:
   goto retry;
 }
 
-static void _upb_Arena_DoFuseArenaLists(upb_ArenaInternal* const parent,
-                                        upb_ArenaInternal* child) {
+// Logically performs the following operation, in a way that is safe against
+// racing fuses:
+//   ret = TAIL(parent)
+//   ret->next = child
+//   return ret
+//
+// The caller is therefore guaranteed that ret->next == child.
+static upb_ArenaInternal* _upb_Arena_LinkForward(
+    upb_ArenaInternal* const parent, upb_ArenaInternal* child) {
   UPB_TSAN_CHECK_PUBLISHED(parent);
   uintptr_t parent_previous_or_tail =
       upb_Atomic_Load(&parent->previous_or_tail, memory_order_acquire);
-  upb_ArenaInternal* parent_tail = parent;
-  if (_upb_Arena_IsTaggedTail(parent_previous_or_tail)) {
-    // Our tail might be stale, but it will always converge to the true tail.
-    parent_tail = _upb_Arena_TailFromTagged(parent_previous_or_tail);
-  }
 
-  // Link parent to child going forwards
-  while (true) {
-    UPB_TSAN_CHECK_PUBLISHED(parent_tail);
-    upb_ArenaInternal* parent_tail_next =
-        upb_Atomic_Load(&parent_tail->next, memory_order_acquire);
+  // Optimization: use parent->previous_or_tail to skip to TAIL(parent) in O(1)
+  // time when possible. This is the common case because we just fused into
+  // parent, suggesting that it should be a root with a cached tail.
+  //
+  // However, if there was a racing fuse, parent may no longer be a root, in
+  // which case we need to walk the entire list to find the tail. The tail
+  // pointer is also not guaranteed to be the true tail, so even when the
+  // optimization is taken, we still need to walk list nodes to find the true
+  // tail.
+  upb_ArenaInternal* parent_tail =
+      _upb_Arena_IsTaggedTail(parent_previous_or_tail)
+          ? _upb_Arena_TailFromTagged(parent_previous_or_tail)
+          : parent;
+
+  UPB_TSAN_CHECK_PUBLISHED(parent_tail);
+  upb_ArenaInternal* parent_tail_next =
+      upb_Atomic_Load(&parent_tail->next, memory_order_acquire);
+
+  do {
+    // Walk the list to find the true tail (a node with next == NULL).
     while (parent_tail_next != NULL) {
       parent_tail = parent_tail_next;
       UPB_TSAN_CHECK_PUBLISHED(parent_tail);
       parent_tail_next =
           upb_Atomic_Load(&parent_tail->next, memory_order_acquire);
     }
-    if (upb_Atomic_CompareExchangeWeak(&parent_tail->next, &parent_tail_next,
-                                       child, memory_order_release,
-                                       memory_order_acquire)) {
-      break;
-    }
-    if (parent_tail_next != NULL) {
-      parent_tail = parent_tail_next;
-    }
-  }
+  } while (!upb_Atomic_CompareExchangeWeak(  // Replace a NULL next with child.
+      &parent_tail->next, &parent_tail_next, child, memory_order_release,
+      memory_order_acquire));
+
+  return parent_tail;
+}
 
-  // Update parent's tail (may be stale).
+// Updates parent->previous_or_tail = child->previous_or_tail in hopes that the
+// latter represents the true tail of the newly-combined list.
+//
+// This is a best-effort operation that may set the tail to a stale value, and
+// may fail to update the tail at all.
+void _upb_Arena_UpdateParentTail(upb_ArenaInternal* parent,
+                                 upb_ArenaInternal* child) {
+  // We are guaranteed that child->previous_or_tail is tagged, because we have
+  // just transitioned child from root -> non-root, which is an exclusive
+  // operation that can only happen once. So we are the exclusive updater of
+  // child->previous_or_tail that can transition it from tagged to untagged.
+  //
+  // However, we are not guaranteed that child->previous_or_tail is the true
+  // tail.  A racing fuse may have appended to child's list but not yet updated
+  // child->previous_or_tail.
   uintptr_t child_previous_or_tail =
       upb_Atomic_Load(&child->previous_or_tail, memory_order_acquire);
   upb_ArenaInternal* new_parent_tail =
       _upb_Arena_TailFromTagged(child_previous_or_tail);
   UPB_TSAN_CHECK_PUBLISHED(new_parent_tail);
 
-  // If another thread fused with us, don't overwrite their previous pointer
-  // with our tail. Relaxed order is fine here as we only inspect the tag bit
-  parent_previous_or_tail =
+  // If another thread fused with parent, such that it is no longer a root,
+  // don't overwrite their previous pointer with our tail. Relaxed order is fine
+  // here as we only inspect the tag bit.
+  uintptr_t parent_previous_or_tail =
       upb_Atomic_Load(&parent->previous_or_tail, memory_order_relaxed);
   if (_upb_Arena_IsTaggedTail(parent_previous_or_tail)) {
     upb_Atomic_CompareExchangeStrong(
@@ -526,13 +614,38 @@ static void _upb_Arena_DoFuseArenaLists(upb_ArenaInternal* const parent,
         _upb_Arena_TaggedFromTail(new_parent_tail), memory_order_release,
         memory_order_relaxed);
   }
+}
 
-  // Link child to parent going backwards, for SpaceAllocated
+static void _upb_Arena_LinkBackward(upb_ArenaInternal* child,
+                                    upb_ArenaInternal* old_parent_tail) {
+  // Link child to parent going backwards, for SpaceAllocated.  This transitions
+  // child->previous_or_tail from tail (tagged) to previous (untagged), after
+  // which its value is immutable.
+  //
+  // - We are guaranteed that no other threads are also attempting to perform
+  //   this transition (tail -> previous), because we just updated
+  //   old_parent_tail->next from NULL to non-NULL, an exclusive operation that
+  //   can only happen once.
+  //
+  // - _upb_Arena_UpdateParentTail() uses CAS to ensure that it
+  //    does not perform the reverse transition (previous -> tail).
+  //
+  // - We are guaranteed that old_parent_tail is the correct "previous" pointer,
+  //   even in the presence of racing fuses that are adding more nodes to the
+  //   list, because _upb_Arena_LinkForward() guarantees that:
+  //       old_parent_tail->next == child.
   upb_Atomic_Store(&child->previous_or_tail,
-                   _upb_Arena_TaggedFromPrevious(parent_tail),
+                   _upb_Arena_TaggedFromPrevious(old_parent_tail),
                    memory_order_release);
 }
 
+static void _upb_Arena_DoFuseArenaLists(upb_ArenaInternal* const parent,
+                                        upb_ArenaInternal* child) {
+  upb_ArenaInternal* old_parent_tail = _upb_Arena_LinkForward(parent, child);
+  _upb_Arena_UpdateParentTail(parent, child);
+  _upb_Arena_LinkBackward(child, old_parent_tail);
+}
+
 void upb_Arena_SetAllocCleanup(upb_Arena* a, upb_AllocCleanupFunc* func) {
   UPB_TSAN_CHECK_READ(a->UPB_ONLYBITS(ptr));
   upb_ArenaInternal* ai = upb_Arena_Internal(a);
@@ -723,3 +836,13 @@ void UPB_PRIVATE(_upb_Arena_SwapOut)(upb_Arena* des, const upb_Arena* src) {
   *des = *src;
   desi->blocks = srci->blocks;
 }
+
+bool _upb_Arena_WasLastAlloc(struct upb_Arena* a, void* ptr, size_t oldsize) {
+  upb_ArenaInternal* ai = upb_Arena_Internal(a);
+  upb_MemBlock* block = ai->blocks;
+  if (block == NULL) return false;
+  block = block->next;
+  if (block == NULL) return false;
+  char* start = UPB_PTR_AT(block, kUpb_MemblockReserve, char);
+  return ptr == start && oldsize == block->size_or_hint - kUpb_MemblockReserve;
+}

data/third_party/upb/upb/mem/arena.h

@@ -110,6 +110,16 @@ void upb_Arena_SetMaxBlockSize(size_t max);
 UPB_API_INLINE void upb_Arena_ShrinkLast(upb_Arena* a, void* ptr,
                                          size_t oldsize, size_t size);
 
+// Attempts to extend the given alloc from arena, in place. Is generally
+// only likely to succeed for the most recent allocation from this arena. If it
+// succeeds, returns true and `ptr`'s allocation is now `size` rather than
+// `oldsize`. Returns false if the allocation cannot be extended; `ptr`'s
+// allocation is unmodified. See also upb_Arena_Realloc.
+// REQUIRES: `size > oldsize`; to shrink, use `upb_Arena_Realloc` or
+// `upb_Arena_ShrinkLast`.
+UPB_API_INLINE bool upb_Arena_TryExtend(upb_Arena* a, void* ptr, size_t oldsize,
+                                        size_t size);
+
 #ifdef UPB_TRACING_ENABLED
 void upb_Arena_SetTraceHandler(void (*initArenaTraceHandler)(const upb_Arena*,
                                                              size_t size),