[Development] Let's get rid of qDebug/qWarning/qCritical!
Koehne Kai
Kai.Koehne at digia.com
Fri Jan 10 13:04:23 CET 2014
> -----Original Message-----
> From: development-bounces+kai.koehne=digia.com at qt-project.org
> [...]
> As a blueprint I've started a patch for a 'qt.core.io' category:
>
> https://codereview.qt-project.org/#change,74862,patchset=1
Andre raised concerns about the (runtime/space) overhead of the QDebug streaming style, compared to the printf style. Here are some conclusions (gcc 4.8, release build, CONFIG+=c++11) ...
RUNTIME OVERHEAD
I did a quick benchmark where I did following variants in a QBENCHMARK (category was enabled, to output was generated to a file. I had to patch qtestlib to not intercept the qWarning output...):
void CatLogBench::qWarningPrintf1() { QBENCHMARK { qWarning("Hi there"); } }
void CatLogBench::qWarningPrintf2() { QBENCHMARK { qWarning("%s", "Hi there"); } }
void CatLogBench::qWarningStream() { QBENCHMARK { qWarning() << "Hi there"; } }
void CatLogBench::qCWarningStream() { QBENCHMARK { qCWarning(cat) << "Hi there"; } }
void CatLogBench::qCWarningPrintf() { QBENCHMARK { qCWarning(cat, "Hi there"); } }
(the last one is a new variant enabled by https://codereview.qt-project.org/#change,75029 .)
And here are the results:
PASS : CatLogBench::qWarningPrintf1()
RESULT : CatLogBench::qWarningPrintf1():
0.0010 msecs per iteration (total: 67, iterations: 65536)
PASS : CatLogBench::qWarningPrintf2()
RESULT : CatLogBench::qWarningPrintf2():
0.0013 msecs per iteration (total: 90, iterations: 65536)
PASS : CatLogBench::qWarningStream()
RESULT : CatLogBench::qWarningStream():
0.0016 msecs per iteration (total: 54, iterations: 32768)
PASS : CatLogBench::qCWarningStream()
RESULT : CatLogBench::qCWarningStream():
0.0018 msecs per iteration (total: 59, iterations: 32768)
PASS : CatLogBench::qCWarningPrintf()
RESULT : CatLogBench::qCWarningPrintf():
0.0010 msecs per iteration (total: 69, iterations: 65536)
PASS : CatLogBench::cleanupTestCase()
Totals: 7 passed, 0 failed, 0 skipped
So, there is of course some measurable overhead, but I wouldn't say it disqualifies the streaming variant per se :)
BINARY SIZE
On to the binary size. I removed the QBENCHMARK macro and disassembled the gcc output. Here it shows that QDebug is completely inlined, the streaming operator variants do indeed generate a hell lot of instructions ! But since this is quite a bogus metric here are the size increase of Qt5Core with, and without the core.io patch included (again a release non-developer build with gcc 4.8.2):
-rwxr-xr-x 1 kkoehne users 5311213 Jan 10 12:55 libQt5Core.so.5.3.0.original
-rwxr-xr-x 1 kkoehne users 5357104 Jan 10 12:58 libQt5Core.so.5.3.0_patch
So the overhead is measurable.
Given these concerns, I'd like to propose adding a printf style overload to qCDebug:
https://codereview.qt-project.org/#change,75029
Note the patch is fairly ugly because we're gracefully handling the case where the toolchain doesn't respect Q_COMPILER_VARIADIC_MACROS, I'll post a second mail about this soon ...
Regards
Kai
More information about the Development
mailing list