[api] Emit raw tag+value writes for forced fixed32 key fields (#15051)

esphome/components/api/proto.h

@@ -236,6 +236,21 @@ class ProtoWriteBuffer {
    * Following https://protobuf.dev/programming-guides/encoding/#structure
    */
   void encode_field_raw(uint32_t field_id, uint32_t type) { this->encode_varint_raw((field_id << 3) | type); }
+  /// Write a precomputed tag byte + 32-bit value in one operation.
+  /// Tag must be a single-byte varint (< 128). No zero check.
+  inline void write_tag_and_fixed32(uint8_t tag, uint32_t value) ESPHOME_ALWAYS_INLINE {
+    this->debug_check_bounds_(5);
+    this->pos_[0] = tag;
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+    std::memcpy(this->pos_ + 1, &value, 4);
+#else
+    this->pos_[1] = static_cast<uint8_t>(value & 0xFF);
+    this->pos_[2] = static_cast<uint8_t>((value >> 8) & 0xFF);
+    this->pos_[3] = static_cast<uint8_t>((value >> 16) & 0xFF);
+    this->pos_[4] = static_cast<uint8_t>((value >> 24) & 0xFF);
+#endif
+    this->pos_ += 5;
+  }
   void encode_string(uint32_t field_id, const char *string, size_t len, bool force = false) {
     if (len == 0 && !force)
       return;
@@ -276,8 +291,7 @@ class ProtoWriteBuffer {
     this->debug_check_bounds_(1);
     *this->pos_++ = value ? 0x01 : 0x00;
   }
-  // noinline: 51 call sites; inlining causes net code growth vs a single out-of-line copy
-  __attribute__((noinline)) void encode_fixed32(uint32_t field_id, uint32_t value, bool force = false) {
+  void encode_fixed32(uint32_t field_id, uint32_t value, bool force = false) {
     if (value == 0 && !force)
       return;
 

script/api_protobuf/api_protobuf.py

@@ -254,14 +254,17 @@ class TypeInfo(ABC):
     def dump(self, name: str) -> str:
         """Dump the value to the output."""
 
+    def calculate_tag(self) -> int:
+        """Calculate the protobuf tag (field_id << 3 | wire_type)."""
+        return (self.number << 3) | (self.wire_type & 0b111)
+
     def calculate_field_id_size(self) -> int:
         """Calculates the size of a field ID in bytes.
 
         Returns:
             The number of bytes needed to encode the field ID
         """
-        # Calculate the tag by combining field_id and wire_type
-        tag = (self.number << 3) | (self.wire_type & 0b111)
+        tag = self.calculate_tag()
 
         # Calculate the varint size
         if tag < 128:
@@ -556,6 +559,16 @@ class Fixed32Type(TypeInfo):
         o += "out.append(buffer);"
         return o
 
+    @property
+    def encode_content(self) -> str:
+        tag = self.calculate_tag()
+        if self.force and tag < 128:
+            # Emit combined tag+value write: precomputed tag + direct memcpy
+            return f"buffer.write_tag_and_fixed32({tag}, this->{self.field_name});"
+        if self.force:
+            return f"buffer.{self.encode_func}({self.number}, this->{self.field_name}, true);"
+        return f"buffer.{self.encode_func}({self.number}, this->{self.field_name});"
+
     def get_size_calculation(self, name: str, force: bool = False) -> str:
         field_id_size = self.calculate_field_id_size()
         if force: