From bd158fc23849663f1cdb430ddea2dbab9deb336f Mon Sep 17 00:00:00 2001 From: Feng Xiao Date: Thu, 10 Nov 2016 16:38:04 -0800 Subject: Speed up JSON parsing. It turns out calling StringOutputStream::Next()/BackUp() repeatedly is very costly in opensource protobuf because it keeps resize() the string back and forth. The current JSON conversion API suffers this problem and leads to ridiculously long parsing time: https://github.com/google/protobuf/issues/2305#issuecomment-257785492 This change fixes the problem but caching the buffer of Next() and avoid calling BackUp() as much as possible. --- src/google/protobuf/util/json_util.cc | 29 ++++++++++++++++------------- 1 file changed, 16 insertions(+), 13 deletions(-) (limited to 'src/google/protobuf/util/json_util.cc') diff --git a/src/google/protobuf/util/json_util.cc b/src/google/protobuf/util/json_util.cc index d7ac2dba..5974cc4a 100644 --- a/src/google/protobuf/util/json_util.cc +++ b/src/google/protobuf/util/json_util.cc @@ -49,22 +49,25 @@ namespace protobuf { namespace util { namespace internal { +ZeroCopyStreamByteSink::~ZeroCopyStreamByteSink() { + stream_->BackUp(buffer_size_); +} + void ZeroCopyStreamByteSink::Append(const char* bytes, size_t len) { - while (len > 0) { - void* buffer; - int length; - if (!stream_->Next(&buffer, &length)) { - // There isn't a way for ByteSink to report errors. + while (true) { + if (len <= buffer_size_) { + memcpy(buffer_, bytes, len); + buffer_ = static_cast(buffer_) + len; + buffer_size_ -= len; return; } - if (len < length) { - memcpy(buffer, bytes, len); - stream_->BackUp(length - len); - break; - } else { - memcpy(buffer, bytes, length); - bytes += length; - len -= length; + memcpy(buffer_, bytes, buffer_size_); + bytes += buffer_size_; + len -= buffer_size_; + if (!stream_->Next(&buffer_, &buffer_size_)) { + // There isn't a way for ByteSink to report errors. + buffer_size_ = 0; + return; } } } -- cgit v1.2.3