diff --git a/cmd/protoc-gen-tarantool/internal/gen/inline.go b/cmd/protoc-gen-tarantool/internal/gen/inline.go
index 17408cb65a6a1c3cbb4590f9eaec2912c8ede496..a14801529e0f901d4faf50072f48dcf0c6a2ffac 100644
--- a/cmd/protoc-gen-tarantool/internal/gen/inline.go
+++ b/cmd/protoc-gen-tarantool/internal/gen/inline.go
@@ -114,10 +114,18 @@ 		emitInlineEncodeMap(w, f, tag, file, selfPath, imports, prefix)
 	case f.Desc.IsList():
 		emitInlineEncodeRepeated(w, f, tag, fname, file, selfPath, imports, prefix)
 	case f.Message != nil:
+		// Split the length-delimited payload into separate `out` slots:
+		// emit the tag, then `varint(#body)`, then the body. This avoids
+		// the per-field `varint(#body) .. body` string concatenation that
+		// `wire.encode_len` would do. `table.concat` at the end of
+		// `_encode` joins everything in one pass — same final bytes,
+		// one fewer allocation per nested-message field.
 		ref := typeRef(file, f.Message.Desc, selfPath, imports, "_encode", prefix)
 		w.line("    if %s then", gate)
+		w.line("        local _b = %s(v)", ref)
 		w.line("        n = n + 1; out[n] = %s", tag)
-		w.line("        n = n + 1; out[n] = wire.encode_len(%s(v))", ref)
+		w.line("        n = n + 1; out[n] = wire.encode_varint(#_b)")
+		w.line("        n = n + 1; out[n] = _b")
 		w.line("    end")
 	case f.Enum != nil:
 		enumLocal := typeRef(file, f.Enum.Desc, selfPath, imports, "", prefix)
@@ -149,8 +157,18 @@ 			w.line("    if v ~= nil and %s then", scalarNotDefaultExpr(st, "v"))
 		} else {
 			w.line("    if %s then", gate)
 		}
-		w.line("        n = n + 1; out[n] = %s", tag)
-		w.line("        n = n + 1; out[n] = wire.encode_%s(v)", st)
+		if st == "string" || st == "bytes" {
+			// Length-delimited scalar: same split rationale as nested
+			// messages above. `encode_string` / `encode_bytes` would
+			// concatenate the length prefix and body — emit them as
+			// separate `out` slots instead and let table.concat join.
+			w.line("        n = n + 1; out[n] = %s", tag)
+			w.line("        n = n + 1; out[n] = wire.encode_varint(#v)")
+			w.line("        n = n + 1; out[n] = v")
+		} else {
+			w.line("        n = n + 1; out[n] = %s", tag)
+			w.line("        n = n + 1; out[n] = wire.encode_%s(v)", st)
+		}
 		w.line("    end")
 	}
 }
@@ -162,8 +180,10 @@ 		ref := typeRef(file, f.Message.Desc, selfPath, imports, "_encode", prefix)
 		w.line("    if v ~= nil and #v > 0 then")
 		w.line("        local _tag = %s", tag)
 		w.line("        for _i = 1, #v do")
+		w.line("            local _b = %s(v[_i])", ref)
 		w.line("            n = n + 1; out[n] = _tag")
-		w.line("            n = n + 1; out[n] = wire.encode_len(%s(v[_i]))", ref)
+		w.line("            n = n + 1; out[n] = wire.encode_varint(#_b)")
+		w.line("            n = n + 1; out[n] = _b")
 		w.line("        end")
 		w.line("    end")
 	case f.Enum != nil:
@@ -180,8 +200,10 @@ 		w.line("                if nv == nil then error(\"unknown enum value '\" .. elem .. \"' for %s\", 0) end", fullName)
 		w.line("            end")
 		w.line("            m = m + 1; parts[m] = wire.encode_int32(nv)")
 		w.line("        end")
+		w.line("        local _b = table.concat(parts)")
 		w.line("        n = n + 1; out[n] = %s", tag)
-		w.line("        n = n + 1; out[n] = wire.encode_len(table.concat(parts))")
+		w.line("        n = n + 1; out[n] = wire.encode_varint(#_b)")
+		w.line("        n = n + 1; out[n] = _b")
 		w.line("    end")
 	default:
 		st := scalarName(f.Desc.Kind())
@@ -192,8 +214,20 @@ 			w.line("        local parts, m = {}, 0")
 			w.line("        for _i = 1, #v do")
 			w.line("            m = m + 1; parts[m] = wire.encode_%s(v[_i])", st)
 			w.line("        end")
+			w.line("        local _b = table.concat(parts)")
 			w.line("        n = n + 1; out[n] = %s", tag)
-			w.line("        n = n + 1; out[n] = wire.encode_len(table.concat(parts))")
+			w.line("        n = n + 1; out[n] = wire.encode_varint(#_b)")
+			w.line("        n = n + 1; out[n] = _b")
+			w.line("    end")
+		} else if st == "string" || st == "bytes" {
+			w.line("    if v ~= nil and #v > 0 then")
+			w.line("        local _tag = %s", tag)
+			w.line("        for _i = 1, #v do")
+			w.line("            local _b = v[_i]")
+			w.line("            n = n + 1; out[n] = _tag")
+			w.line("            n = n + 1; out[n] = wire.encode_varint(#_b)")
+			w.line("            n = n + 1; out[n] = _b")
+			w.line("        end")
 			w.line("    end")
 		} else {
 			w.line("    if v ~= nil and #v > 0 then")
diff --git a/examples/expected/full/conformance/conformance_pb.lua b/examples/expected/full/conformance/conformance_pb.lua
index 150c6b8d088b0475e2e569c044741d6543f4e870..12439fb41ac9f9dd9caca5745154b08d6dba85b0 100644
--- a/examples/expected/full/conformance/conformance_pb.lua
+++ b/examples/expected/full/conformance/conformance_pb.lua
@@ -145,19 +145,22 @@     -- field 1: name
     v = t.name
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 2: failure_message
     v = t.failure_message
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x12"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 3: matched_name
     v = t.matched_name
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x1a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
@@ -224,8 +227,10 @@     v = t.test
     if v ~= nil and #v > 0 then
         local _tag = "\x12"
         for _i = 1, #v do
+            local _b = M.TestStatus_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(M.TestStatus_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     local _uf = t._unknown_fields
@@ -291,25 +296,29 @@     -- field 1: protobuf_payload
     v = t.protobuf_payload
     if _of_payload == "protobuf_payload" then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_bytes(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 2: json_payload
     v = t.json_payload
     if _of_payload == "json_payload" then
         n = n + 1; out[n] = "\x12"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 7: jspb_payload
     v = t.jspb_payload
     if _of_payload == "jspb_payload" then
         n = n + 1; out[n] = "\x3a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 8: text_payload
     v = t.text_payload
     if _of_payload == "text_payload" then
         n = n + 1; out[n] = "\x42"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 3: requested_output_format
     v = t.requested_output_format
@@ -328,7 +337,8 @@     -- field 4: message_type
     v = t.message_type
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x22"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 5: test_category
     v = t.test_category
@@ -346,8 +356,10 @@     end
     -- field 6: jspb_encoding_options
     v = t.jspb_encoding_options
     if v ~= nil or type(v) == 'cdata' then
+        local _b = M.JspbEncodingConfig_encode(v)
         n = n + 1; out[n] = "\x32"
-        n = n + 1; out[n] = wire.encode_len(M.JspbEncodingConfig_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 9: print_unknown_fields
     v = t.print_unknown_fields
@@ -470,55 +482,64 @@     -- field 1: parse_error
     v = t.parse_error
     if _of_result == "parse_error" then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 6: serialize_error
     v = t.serialize_error
     if _of_result == "serialize_error" then
         n = n + 1; out[n] = "\x32"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 9: timeout_error
     v = t.timeout_error
     if _of_result == "timeout_error" then
         n = n + 1; out[n] = "\x4a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 2: runtime_error
     v = t.runtime_error
     if _of_result == "runtime_error" then
         n = n + 1; out[n] = "\x12"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 3: protobuf_payload
     v = t.protobuf_payload
     if _of_result == "protobuf_payload" then
         n = n + 1; out[n] = "\x1a"
-        n = n + 1; out[n] = wire.encode_bytes(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 4: json_payload
     v = t.json_payload
     if _of_result == "json_payload" then
         n = n + 1; out[n] = "\x22"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 5: skipped
     v = t.skipped
     if _of_result == "skipped" then
         n = n + 1; out[n] = "\x2a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 7: jspb_payload
     v = t.jspb_payload
     if _of_result == "jspb_payload" then
         n = n + 1; out[n] = "\x3a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 8: text_payload
     v = t.text_payload
     if _of_result == "text_payload" then
         n = n + 1; out[n] = "\x42"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
diff --git a/examples/expected/full/hello/hello_pb.lua b/examples/expected/full/hello/hello_pb.lua
index 0bfef1adae3efafd1a829d4f21bd5afe406671b6..1ac6a88a2d490ea66a5dc6dba6791ab0871dbb95 100644
--- a/examples/expected/full/hello/hello_pb.lua
+++ b/examples/expected/full/hello/hello_pb.lua
@@ -172,7 +172,8 @@     -- field 2: text
     v = t.text
     if _of_outcome == "text" then
         n = n + 1; out[n] = "\x12"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 3: code
     v = t.code
@@ -183,8 +184,10 @@     end
     -- field 4: details
     v = t.details
     if _of_outcome == "details" then
+        local _b = M.Address_encode(v)
         n = n + 1; out[n] = "\x22"
-        n = n + 1; out[n] = wire.encode_len(M.Address_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
@@ -265,7 +268,8 @@     -- field 1: name
     v = t.name
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
@@ -323,7 +327,8 @@     -- field 1: greeting
     v = t.greeting
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
@@ -381,73 +386,96 @@     -- field 1: title
     v = t.title
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 2: created_at
     v = t.created_at
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Timestamp_encode(v)
         n = n + 1; out[n] = "\x12"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Timestamp_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 3: duration
     v = t.duration
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Duration_encode(v)
         n = n + 1; out[n] = "\x1a"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Duration_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 4: ack
     v = t.ack
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Empty_encode(v)
         n = n + 1; out[n] = "\x22"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Empty_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 5: retry_count
     v = t.retry_count
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Int32Value_encode(v)
         n = n + 1; out[n] = "\x2a"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Int32Value_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 6: note
     v = t.note
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.StringValue_encode(v)
         n = n + 1; out[n] = "\x32"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.StringValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 7: is_admin
     v = t.is_admin
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.BoolValue_encode(v)
         n = n + 1; out[n] = "\x3a"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.BoolValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 8: payload
     v = t.payload
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Struct_encode(v)
         n = n + 1; out[n] = "\x42"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Struct_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 9: attribute
     v = t.attribute
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Value_encode(v)
         n = n + 1; out[n] = "\x4a"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Value_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 10: tags
     v = t.tags
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.ListValue_encode(v)
         n = n + 1; out[n] = "\x52"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.ListValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 11: extension
     v = t.extension
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Any_encode(v)
         n = n + 1; out[n] = "\x5a"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Any_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 12: update_mask
     v = t.update_mask
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.FieldMask_encode(v)
         n = n + 1; out[n] = "\x62"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.FieldMask_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
@@ -549,13 +577,15 @@     -- field 1: street
     v = t.street
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 2: city
     v = t.city
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x12"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 3: zip
     v = t.zip
@@ -567,7 +597,8 @@     -- field 4: apartment
     v = t.apartment
     if v ~= nil then
         n = n + 1; out[n] = "\x22"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
@@ -642,7 +673,8 @@     -- field 1: name
     v = t.name
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x0a"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 2: age
     v = t.age
@@ -655,8 +687,10 @@     v = t.emails
     if v ~= nil and #v > 0 then
         local _tag = "\x1a"
         for _i = 1, #v do
+            local _b = v[_i]
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_string(v[_i])
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 4: status
@@ -675,16 +709,20 @@     end
     -- field 5: address
     v = t.address
     if v ~= nil or type(v) == 'cdata' then
+        local _b = M.Address_encode(v)
         n = n + 1; out[n] = "\x2a"
-        n = n + 1; out[n] = wire.encode_len(M.Address_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 6: friends
     v = t.friends
     if v ~= nil and #v > 0 then
         local _tag = "\x32"
         for _i = 1, #v do
+            local _b = M.Person_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(M.Person_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 7: lucky_numbers
@@ -694,14 +732,17 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_int32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x3a"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 8: avatar
     v = t.avatar
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x42"
-        n = n + 1; out[n] = wire.encode_bytes(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 9: user_id
     v = t.user_id
diff --git a/examples/expected/full/protobuf_test_messages/proto3/test_messages_proto3_pb.lua b/examples/expected/full/protobuf_test_messages/proto3/test_messages_proto3_pb.lua
index 18c58a2d178b1c30849435b8805b7a6c83f37e91..759e89fa99d95fe5c4e866c30d811a0275d63fde 100644
--- a/examples/expected/full/protobuf_test_messages/proto3/test_messages_proto3_pb.lua
+++ b/examples/expected/full/protobuf_test_messages/proto3/test_messages_proto3_pb.lua
@@ -513,25 +513,31 @@     -- field 14: optional_string
     v = t.optional_string
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x72"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 15: optional_bytes
     v = t.optional_bytes
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\x7a"
-        n = n + 1; out[n] = wire.encode_bytes(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 18: optional_nested_message
     v = t.optional_nested_message
     if v ~= nil or type(v) == 'cdata' then
+        local _b = M.TestAllTypesProto3_NestedMessage_encode(v)
         n = n + 1; out[n] = "\x92\x01"
-        n = n + 1; out[n] = wire.encode_len(M.TestAllTypesProto3_NestedMessage_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 19: optional_foreign_message
     v = t.optional_foreign_message
     if v ~= nil or type(v) == 'cdata' then
+        local _b = M.ForeignMessage_encode(v)
         n = n + 1; out[n] = "\x9a\x01"
-        n = n + 1; out[n] = wire.encode_len(M.ForeignMessage_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 21: optional_nested_enum
     v = t.optional_nested_enum
@@ -576,19 +582,23 @@     -- field 24: optional_string_piece
     v = t.optional_string_piece
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\xc2\x01"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 25: optional_cord
     v = t.optional_cord
     if v ~= nil and v ~= '' then
         n = n + 1; out[n] = "\xca\x01"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 27: recursive_message
     v = t.recursive_message
     if v ~= nil or type(v) == 'cdata' then
+        local _b = M.TestAllTypesProto3_encode(v)
         n = n + 1; out[n] = "\xda\x01"
-        n = n + 1; out[n] = wire.encode_len(M.TestAllTypesProto3_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 31: repeated_int32
     v = t.repeated_int32
@@ -597,8 +607,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_int32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xfa\x01"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 32: repeated_int64
     v = t.repeated_int64
@@ -607,8 +619,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_int64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x82\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 33: repeated_uint32
     v = t.repeated_uint32
@@ -617,8 +631,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_uint32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x8a\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 34: repeated_uint64
     v = t.repeated_uint64
@@ -627,8 +643,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_uint64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x92\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 35: repeated_sint32
     v = t.repeated_sint32
@@ -637,8 +655,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sint32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x9a\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 36: repeated_sint64
     v = t.repeated_sint64
@@ -647,8 +667,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sint64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xa2\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 37: repeated_fixed32
     v = t.repeated_fixed32
@@ -657,8 +679,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_fixed32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xaa\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 38: repeated_fixed64
     v = t.repeated_fixed64
@@ -667,8 +691,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_fixed64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xb2\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 39: repeated_sfixed32
     v = t.repeated_sfixed32
@@ -677,8 +703,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sfixed32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xba\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 40: repeated_sfixed64
     v = t.repeated_sfixed64
@@ -687,8 +715,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sfixed64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xc2\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 41: repeated_float
     v = t.repeated_float
@@ -697,8 +727,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_float(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xca\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 42: repeated_double
     v = t.repeated_double
@@ -707,8 +739,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_double(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xd2\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 43: repeated_bool
     v = t.repeated_bool
@@ -717,16 +751,20 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_bool(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xda\x02"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 44: repeated_string
     v = t.repeated_string
     if v ~= nil and #v > 0 then
         local _tag = "\xe2\x02"
         for _i = 1, #v do
+            local _b = v[_i]
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_string(v[_i])
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 45: repeated_bytes
@@ -734,8 +772,10 @@     v = t.repeated_bytes
     if v ~= nil and #v > 0 then
         local _tag = "\xea\x02"
         for _i = 1, #v do
+            local _b = v[_i]
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_bytes(v[_i])
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 48: repeated_nested_message
@@ -743,8 +783,10 @@     v = t.repeated_nested_message
     if v ~= nil and #v > 0 then
         local _tag = "\x82\x03"
         for _i = 1, #v do
+            local _b = M.TestAllTypesProto3_NestedMessage_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(M.TestAllTypesProto3_NestedMessage_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 49: repeated_foreign_message
@@ -752,8 +794,10 @@     v = t.repeated_foreign_message
     if v ~= nil and #v > 0 then
         local _tag = "\x8a\x03"
         for _i = 1, #v do
+            local _b = M.ForeignMessage_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(M.ForeignMessage_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 51: repeated_nested_enum
@@ -769,8 +813,10 @@                 if nv == nil then error("unknown enum value '" .. elem .. "' for protobuf_test_messages.proto3.TestAllTypesProto3.NestedEnum", 0) end
             end
             m = m + 1; parts[m] = wire.encode_int32(nv)
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x9a\x03"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 52: repeated_foreign_enum
     v = t.repeated_foreign_enum
@@ -785,16 +831,20 @@                 if nv == nil then error("unknown enum value '" .. elem .. "' for protobuf_test_messages.proto3.ForeignEnum", 0) end
             end
             m = m + 1; parts[m] = wire.encode_int32(nv)
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xa2\x03"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 54: repeated_string_piece
     v = t.repeated_string_piece
     if v ~= nil and #v > 0 then
         local _tag = "\xb2\x03"
         for _i = 1, #v do
+            local _b = v[_i]
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_string(v[_i])
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 55: repeated_cord
@@ -802,8 +852,10 @@     v = t.repeated_cord
     if v ~= nil and #v > 0 then
         local _tag = "\xba\x03"
         for _i = 1, #v do
+            local _b = v[_i]
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_string(v[_i])
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 75: packed_int32
@@ -813,8 +865,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_int32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xda\x04"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 76: packed_int64
     v = t.packed_int64
@@ -823,8 +877,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_int64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xe2\x04"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 77: packed_uint32
     v = t.packed_uint32
@@ -833,8 +889,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_uint32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xea\x04"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 78: packed_uint64
     v = t.packed_uint64
@@ -843,8 +901,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_uint64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xf2\x04"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 79: packed_sint32
     v = t.packed_sint32
@@ -853,8 +913,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sint32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xfa\x04"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 80: packed_sint64
     v = t.packed_sint64
@@ -863,8 +925,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sint64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x82\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 81: packed_fixed32
     v = t.packed_fixed32
@@ -873,8 +937,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_fixed32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x8a\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 82: packed_fixed64
     v = t.packed_fixed64
@@ -883,8 +949,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_fixed64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x92\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 83: packed_sfixed32
     v = t.packed_sfixed32
@@ -893,8 +961,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sfixed32(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\x9a\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 84: packed_sfixed64
     v = t.packed_sfixed64
@@ -903,8 +973,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_sfixed64(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xa2\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 85: packed_float
     v = t.packed_float
@@ -913,8 +985,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_float(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xaa\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 86: packed_double
     v = t.packed_double
@@ -923,8 +997,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_double(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xb2\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 87: packed_bool
     v = t.packed_bool
@@ -933,8 +1009,10 @@         local parts, m = {}, 0
         for _i = 1, #v do
             m = m + 1; parts[m] = wire.encode_bool(v[_i])
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xba\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 88: packed_nested_enum
     v = t.packed_nested_enum
@@ -949,8 +1027,10 @@                 if nv == nil then error("unknown enum value '" .. elem .. "' for protobuf_test_messages.proto3.TestAllTypesProto3.NestedEnum", 0) end
             end
             m = m + 1; parts[m] = wire.encode_int32(nv)
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xc2\x05"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 89: unpacked_int32
     v = t.unpacked_int32
@@ -1082,8 +1162,10 @@                 if nv == nil then error("unknown enum value '" .. elem .. "' for protobuf_test_messages.proto3.TestAllTypesProto3.NestedEnum", 0) end
             end
             m = m + 1; parts[m] = wire.encode_int32(nv)
         end
+        local _b = table.concat(parts)
         n = n + 1; out[n] = "\xb0\x06"
-        n = n + 1; out[n] = wire.encode_len(table.concat(parts))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 56: map_int32_int32
     v = t.map_int32_int32
@@ -1446,20 +1528,24 @@     end
     -- field 112: oneof_nested_message
     v = t.oneof_nested_message
     if _of_oneof_field == "oneof_nested_message" then
+        local _b = M.TestAllTypesProto3_NestedMessage_encode(v)
         n = n + 1; out[n] = "\x82\x07"
-        n = n + 1; out[n] = wire.encode_len(M.TestAllTypesProto3_NestedMessage_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 113: oneof_string
     v = t.oneof_string
     if _of_oneof_field == "oneof_string" then
         n = n + 1; out[n] = "\x8a\x07"
-        n = n + 1; out[n] = wire.encode_string(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 114: oneof_bytes
     v = t.oneof_bytes
     if _of_oneof_field == "oneof_bytes" then
         n = n + 1; out[n] = "\x92\x07"
-        n = n + 1; out[n] = wire.encode_bytes(v)
+        n = n + 1; out[n] = wire.encode_varint(#v)
+        n = n + 1; out[n] = v
     end
     -- field 115: oneof_bool
     v = t.oneof_bool
@@ -1510,64 +1596,84 @@     end
     -- field 201: optional_bool_wrapper
     v = t.optional_bool_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.BoolValue_encode(v)
         n = n + 1; out[n] = "\xca\x0c"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.BoolValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 202: optional_int32_wrapper
     v = t.optional_int32_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Int32Value_encode(v)
         n = n + 1; out[n] = "\xd2\x0c"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Int32Value_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 203: optional_int64_wrapper
     v = t.optional_int64_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Int64Value_encode(v)
         n = n + 1; out[n] = "\xda\x0c"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Int64Value_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 204: optional_uint32_wrapper
     v = t.optional_uint32_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.UInt32Value_encode(v)
         n = n + 1; out[n] = "\xe2\x0c"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.UInt32Value_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 205: optional_uint64_wrapper
     v = t.optional_uint64_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.UInt64Value_encode(v)
         n = n + 1; out[n] = "\xea\x0c"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.UInt64Value_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 206: optional_float_wrapper
     v = t.optional_float_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.FloatValue_encode(v)
         n = n + 1; out[n] = "\xf2\x0c"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.FloatValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 207: optional_double_wrapper
     v = t.optional_double_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.DoubleValue_encode(v)
         n = n + 1; out[n] = "\xfa\x0c"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.DoubleValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 208: optional_string_wrapper
     v = t.optional_string_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.StringValue_encode(v)
         n = n + 1; out[n] = "\x82\x0d"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.StringValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 209: optional_bytes_wrapper
     v = t.optional_bytes_wrapper
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.BytesValue_encode(v)
         n = n + 1; out[n] = "\x8a\x0d"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.BytesValue_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 211: repeated_bool_wrapper
     v = t.repeated_bool_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\x9a\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.BoolValue_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.BoolValue_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 212: repeated_int32_wrapper
@@ -1575,8 +1681,10 @@     v = t.repeated_int32_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xa2\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.Int32Value_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Int32Value_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 213: repeated_int64_wrapper
@@ -1584,8 +1692,10 @@     v = t.repeated_int64_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xaa\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.Int64Value_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Int64Value_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 214: repeated_uint32_wrapper
@@ -1593,8 +1703,10 @@     v = t.repeated_uint32_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xb2\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.UInt32Value_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.UInt32Value_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 215: repeated_uint64_wrapper
@@ -1602,8 +1714,10 @@     v = t.repeated_uint64_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xba\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.UInt64Value_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.UInt64Value_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 216: repeated_float_wrapper
@@ -1611,8 +1725,10 @@     v = t.repeated_float_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xc2\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.FloatValue_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.FloatValue_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 217: repeated_double_wrapper
@@ -1620,8 +1736,10 @@     v = t.repeated_double_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xca\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.DoubleValue_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.DoubleValue_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 218: repeated_string_wrapper
@@ -1629,8 +1747,10 @@     v = t.repeated_string_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xd2\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.StringValue_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.StringValue_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 219: repeated_bytes_wrapper
@@ -1638,45 +1758,59 @@     v = t.repeated_bytes_wrapper
     if v ~= nil and #v > 0 then
         local _tag = "\xda\x0d"
         for _i = 1, #v do
+            local _b = pb.wkt.BytesValue_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.BytesValue_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 301: optional_duration
     v = t.optional_duration
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Duration_encode(v)
         n = n + 1; out[n] = "\xea\x12"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Duration_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 302: optional_timestamp
     v = t.optional_timestamp
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Timestamp_encode(v)
         n = n + 1; out[n] = "\xf2\x12"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Timestamp_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 303: optional_field_mask
     v = t.optional_field_mask
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.FieldMask_encode(v)
         n = n + 1; out[n] = "\xfa\x12"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.FieldMask_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 304: optional_struct
     v = t.optional_struct
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Struct_encode(v)
         n = n + 1; out[n] = "\x82\x13"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Struct_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 305: optional_any
     v = t.optional_any
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Any_encode(v)
         n = n + 1; out[n] = "\x8a\x13"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Any_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 306: optional_value
     v = t.optional_value
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Value_encode(v)
         n = n + 1; out[n] = "\x92\x13"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Value_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 307: optional_null_value
     v = t.optional_null_value
@@ -1694,16 +1828,20 @@     end
     -- field 308: optional_empty
     v = t.optional_empty
     if v ~= nil or type(v) == 'cdata' then
+        local _b = pb.wkt.Empty_encode(v)
         n = n + 1; out[n] = "\xa2\x13"
-        n = n + 1; out[n] = wire.encode_len(pb.wkt.Empty_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     -- field 311: repeated_duration
     v = t.repeated_duration
     if v ~= nil and #v > 0 then
         local _tag = "\xba\x13"
         for _i = 1, #v do
+            local _b = pb.wkt.Duration_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Duration_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 312: repeated_timestamp
@@ -1711,8 +1849,10 @@     v = t.repeated_timestamp
     if v ~= nil and #v > 0 then
         local _tag = "\xc2\x13"
         for _i = 1, #v do
+            local _b = pb.wkt.Timestamp_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Timestamp_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 313: repeated_fieldmask
@@ -1720,8 +1860,10 @@     v = t.repeated_fieldmask
     if v ~= nil and #v > 0 then
         local _tag = "\xca\x13"
         for _i = 1, #v do
+            local _b = pb.wkt.FieldMask_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.FieldMask_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 324: repeated_struct
@@ -1729,8 +1871,10 @@     v = t.repeated_struct
     if v ~= nil and #v > 0 then
         local _tag = "\xa2\x14"
         for _i = 1, #v do
+            local _b = pb.wkt.Struct_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Struct_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 315: repeated_any
@@ -1738,8 +1882,10 @@     v = t.repeated_any
     if v ~= nil and #v > 0 then
         local _tag = "\xda\x13"
         for _i = 1, #v do
+            local _b = pb.wkt.Any_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Any_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 316: repeated_value
@@ -1747,8 +1893,10 @@     v = t.repeated_value
     if v ~= nil and #v > 0 then
         local _tag = "\xe2\x13"
         for _i = 1, #v do
+            local _b = pb.wkt.Value_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Value_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 317: repeated_list_value
@@ -1756,8 +1904,10 @@     v = t.repeated_list_value
     if v ~= nil and #v > 0 then
         local _tag = "\xea\x13"
         for _i = 1, #v do
+            local _b = pb.wkt.ListValue_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.ListValue_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 318: repeated_empty
@@ -1765,8 +1915,10 @@     v = t.repeated_empty
     if v ~= nil and #v > 0 then
         local _tag = "\xf2\x13"
         for _i = 1, #v do
+            local _b = pb.wkt.Empty_encode(v[_i])
             n = n + 1; out[n] = _tag
-            n = n + 1; out[n] = wire.encode_len(pb.wkt.Empty_encode(v[_i]))
+            n = n + 1; out[n] = wire.encode_varint(#_b)
+            n = n + 1; out[n] = _b
         end
     end
     -- field 401: fieldname1
@@ -3569,8 +3721,10 @@     end
     -- field 2: corecursive
     v = t.corecursive
     if v ~= nil or type(v) == 'cdata' then
+        local _b = M.TestAllTypesProto3_encode(v)
         n = n + 1; out[n] = "\x12"
-        n = n + 1; out[n] = wire.encode_len(M.TestAllTypesProto3_encode(v))
+        n = n + 1; out[n] = wire.encode_varint(#_b)
+        n = n + 1; out[n] = _b
     end
     local _uf = t._unknown_fields
     if _uf ~= nil and _uf ~= '' then n = n + 1; out[n] = _uf end
diff --git a/runtime/pb/codec.lua b/runtime/pb/codec.lua
index d0d635477b885eae7d7b7e49d9088009008c6267..3bc4eb58fd561b29eca7b3fa63abaa3b03939cad 100644
--- a/runtime/pb/codec.lua
+++ b/runtime/pb/codec.lua
@@ -271,6 +271,12 @@         local handler = scalar[f.proto_type]
         if not handler then return nil end
         local encode_value = handler.encode
 
+        -- Split length-prefix emission: emit `tag → varint(#payload)
+        -- → payload` as three separate `out` slots instead of letting
+        -- `wire.encode_len` concatenate the prefix and body. Saves the
+        -- per-field string allocation; table.concat at the end joins
+        -- everything in one pass. Mirror of the inline.go codegen
+        -- change for full mode.
         if f.packed and handler.packable then
             local tag_bytes = wire.encode_tag(f.id, wire.WIRE_LEN)
             return function(data, out)
@@ -283,12 +289,32 @@                 for i = 1, nv do parts[i] = encode_value(v[i]) end
                 local payload = table.concat(parts)
                 local n = #out
                 out[n + 1] = tag_bytes
-                out[n + 2] = wire.encode_len(payload)
+                out[n + 2] = wire.encode_varint(#payload)
+                out[n + 3] = payload
+            end
+        end
+
+        -- Unpacked repeated string/bytes: split each element's length
+        -- prefix from its body, same rationale as above.
+        if f.proto_type == 'string' or f.proto_type == 'bytes' then
+            local tag_bytes = wire.encode_tag(f.id, handler.wire)
+            return function(data, out)
+                local v = data[fname]
+                if v == nil then return end
+                local nv = #v
+                if nv == 0 then return end
+                local n = #out
+                for i = 1, nv do
+                    local s = v[i]
+                    n = n + 1; out[n] = tag_bytes
+                    n = n + 1; out[n] = wire.encode_varint(#s)
+                    n = n + 1; out[n] = s
+                end
             end
         end
 
-        -- Unpacked repeated scalar (also covers string/bytes — LEN
-        -- wire type, which is non-packable by proto3 rules).
+        -- Unpacked repeated numeric/bool scalar: not length-delimited,
+        -- emit tag + encoded value per element.
         local tag_bytes = wire.encode_tag(f.id, handler.wire)
         return function(data, out)
             local v = data[fname]
@@ -313,8 +339,10 @@             local nv = #v
             if nv == 0 then return end
             local n = #out
             for i = 1, nv do
+                local body = encode_msg(sub_desc, v[i])
                 n = n + 1; out[n] = tag_bytes
-                n = n + 1; out[n] = wire.encode_len(encode_msg(sub_desc, v[i]))
+                n = n + 1; out[n] = wire.encode_varint(#body)
+                n = n + 1; out[n] = body
             end
         end
     end
@@ -333,9 +361,11 @@                 local parts = {}
                 for i = 1, nv do
                     parts[i] = wire.encode_varint(encode_enum_value(enum_desc, v[i]))
                 end
+                local payload = table.concat(parts)
                 local n = #out
                 out[n + 1] = tag_bytes
-                out[n + 2] = wire.encode_len(table.concat(parts))
+                out[n + 2] = wire.encode_varint(#payload)
+                out[n + 3] = payload
             end
         end
         local tag_bytes = wire.encode_tag(f.id, wire.WIRE_VARINT)
@@ -383,12 +413,17 @@             end
         end
 
         if proto_type == 'string' or proto_type == 'bytes' then
+            -- Split: emit tag + varint(#v) + v as three separate `out`
+            -- slots, avoiding the per-field `varint(#v) .. v` concat
+            -- that `encode_value` (== wire.encode_string/bytes) would
+            -- do. See packed/repeated writers above for the rationale.
             return function(data, out)
                 local v = data[fname]
                 if v == nil or v == '' then return end
                 local n = #out
                 out[n + 1] = tag_bytes
-                out[n + 2] = encode_value(v)
+                out[n + 2] = wire.encode_varint(#v)
+                out[n + 3] = v
             end
         end
         if proto_type == 'bool' then
@@ -438,9 +473,11 @@             -- google.protobuf.Value uses box.NULL as the canonical
             -- null_value sentinel — its sub_desc.encode handles it. Only
             -- treat the field as absent if it's actually nil (no cdata).
             if v == nil and type(v) ~= 'cdata' then return end
+            local body = encode_msg(sub_desc, v)
             local n = #out
             out[n + 1] = tag_bytes
-            out[n + 2] = wire.encode_len(encode_msg(sub_desc, v))
+            out[n + 2] = wire.encode_varint(#body)
+            out[n + 3] = body
         end
     end
 
diff --git a/runtime/pb/wire.lua b/runtime/pb/wire.lua
index 40eedb77ce6546a93b6dba614a59cca094ceea2a..6a9f58a49d67483a7a4719ddd01d443ebc435c80 100644
--- a/runtime/pb/wire.lua
+++ b/runtime/pb/wire.lua
@@ -84,14 +84,59 @@ -- string.char(n) call with no cdata allocation, no `out` table, no
 -- table.concat. Covers most length prefixes for short strings, many
 -- enum ordinals, and most small int values in typical RPC payloads.
 --
--- We deliberately do NOT extend the fast path to 2-4 byte values:
--- growing the function past the LuaJIT inline budget makes parent
--- traces stop inlining it, which costs more (~30% bench regression
--- on 1-byte-dominant workloads) than the rare multi-byte case gains.
-local function encode_varint(n)
-    if type(n) == 'number' and n >= 0 and n < 0x80 then
-        return string.char(n)
+-- The outer function is kept tiny on purpose — adding 2/3/4-byte
+-- branches to `encode_varint` itself pushes it past LuaJIT's inline
+-- budget so parent traces stop inlining it, regressing the
+-- 1-byte-dominant workloads by ~30%. Multi-byte Lua-number fast paths
+-- live in `encode_varint_slow` below, which is not inlined into hot
+-- traces anyway — so its size doesn't matter.
+local encode_varint_slow
+
+-- encode_varint_slow: handles every input that doesn't fit the
+-- 1-byte fast path in `encode_varint`. Three layers, falling through:
+--
+--   1. Non-negative Lua numbers < 2^28 are encoded via `bit.rshift` /
+--      `bit.band` / `string.char` — bit ops on Lua numbers operate as
+--      uint32 and don't allocate cdata. Values up to 4 bytes covered.
+--   2. Non-negative Lua numbers in [2^28, 2^53) are still exact in
+--      double precision but exceed uint32 — emit one byte through bit
+--      ops, then recurse with `n / 128` on the smaller residue.
+--   3. Everything else (cdata uint64/int64, negative Lua numbers — the
+--      latter get sign-extended to 10-byte varints) goes through the
+--      uint64 cdata loop. Pre-change measurement: 18x slower than the
+--      1-byte fast path; the Lua-number paths above bring 2-byte to
+--      ~2x of the fast path instead.
+encode_varint_slow = function(n)
+    if type(n) == 'number' and n >= 0 then
+        if n < 0x4000 then          -- 2-byte
+            return string.char(
+                bit.bor(bit.band(n, 0x7f), 0x80),
+                bit.rshift(n, 7))
+        end
+        if n < 0x200000 then        -- 3-byte
+            return string.char(
+                bit.bor(bit.band(n, 0x7f), 0x80),
+                bit.bor(bit.band(bit.rshift(n, 7), 0x7f), 0x80),
+                bit.rshift(n, 14))
+        end
+        if n < 0x10000000 then      -- 4-byte (< 2^28, fits in uint32)
+            return string.char(
+                bit.bor(bit.band(n, 0x7f), 0x80),
+                bit.bor(bit.band(bit.rshift(n, 7), 0x7f), 0x80),
+                bit.bor(bit.band(bit.rshift(n, 14), 0x7f), 0x80),
+                bit.rshift(n, 21))
+        end
+        -- 2^28 <= n < 2^53: still exact in double, but past uint32.
+        -- Emit one byte via Lua-number bit op, then recurse on the
+        -- residue (which is now < n/128 and may now fit a fast path).
+        if n < 2^53 then
+            return string.char(bit.bor(bit.band(n, 0x7f), 0x80))
+                .. encode_varint_slow(math.floor(n / 128))
+        end
     end
+    -- Fallback: uint64 cdata path. Reached by cdata inputs, negative
+    -- Lua numbers (sign-extended to 10-byte varint), and the (rare)
+    -- case of a Lua number larger than 2^53.
     n = to_uint64(n)
     local out = {}
     local i = 1
@@ -102,6 +147,13 @@         i = i + 1
     end
     out[i] = string.char(tonumber(n))
     return table.concat(out)
+end
+
+local function encode_varint(n)
+    if type(n) == 'number' and n >= 0 and n < 0x80 then
+        return string.char(n)
+    end
+    return encode_varint_slow(n)
 end
 M.encode_varint = encode_varint