diff options
| author | Marc Mutz <marc.mutz@qt.io> | 2024-12-20 22:00:32 +0100 |
|---|---|---|
| committer | Marc Mutz <marc.mutz@qt.io> | 2024-12-24 23:41:58 +0000 |
| commit | 1a9f8cc0df33195df959cee2e355dde4cbacd754 (patch) | |
| tree | 5320e34e82ac49261260786f789032bd662a93ac /src/corelib/serialization/qdatastream.cpp | |
| parent | 05b9a4b2deefd586356e1f36d84372b06e74cfe3 (diff) | |
Fix a performance regression in QDataStream
Commit e176dd78fd2f253eb2625585b2bd90b5713e5984 replaced a `new
char[n]` with a std::make_unique<char[]>(n), probably on this author's
insistence.
But the two are not equivalent: make_unique() value-initializes, even
arrays, even of built-in type, which means that each buffer resize
writes each byte twice: first nulling out the whole buffer as part of
value-initialization, then in the memcpy() and any following read()s.
For buffers of several MiB, or even GiB in size, this is very costly.
Fix by adding and using a backport of C++20
make_unique_for_overwrite(), which performs the equivalent of the old
code (ie. default-, not value-initialization).
Also add q20::is_(un)bounded_array, which are needed for the
implementation of q20::make_unique_for_overwrite().
Amends e176dd78fd2f253eb2625585b2bd90b5713e5984.
Pick-to: 6.9 6.8 6.5
Change-Id: I8865c7369e522ec475df122e3d00d6aba3b24561
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Diffstat (limited to 'src/corelib/serialization/qdatastream.cpp')
| -rw-r--r-- | src/corelib/serialization/qdatastream.cpp | 4 |
1 files changed, 3 insertions, 1 deletions
diff --git a/src/corelib/serialization/qdatastream.cpp b/src/corelib/serialization/qdatastream.cpp index 039c8fbcdc3..f9c08cd78e4 100644 --- a/src/corelib/serialization/qdatastream.cpp +++ b/src/corelib/serialization/qdatastream.cpp @@ -13,6 +13,8 @@ #include <stdlib.h> #include "qendian.h" +#include <QtCore/q20memory.h> + QT_BEGIN_NAMESPACE constexpr quint32 QDataStream::NullCode; @@ -1096,7 +1098,7 @@ QDataStream &QDataStream::readBytes(char *&s, qint64 &l) do { qsizetype blockSize = qMin(step, len - allocated); const qsizetype n = allocated + blockSize + 1; - if (const auto prevBuf = std::exchange(curBuf, std::make_unique<char[]>(n))) + if (const auto prevBuf = std::exchange(curBuf, q20::make_unique_for_overwrite<char[]>(n))) memcpy(curBuf.get(), prevBuf.get(), allocated); if (readBlock(curBuf.get() + allocated, blockSize) != blockSize) return *this; |
