Changeset 3d003a7a1a87 change 'configwith' behavior so that the default value is
run through the conversion function. In parallel a new user of 'configwith' got
introduced unaware of this coming behavior change. This broke profiling.
We resolve the situation by having the new conversion function cope with a
default value already using the right type.
statprof.display_hotpath() accepts a "limit" function to choose the
minimum threshold for samples to display. The default is 0.05, which
means you don't need individual items contributing less than 5%.
I had a need to adjust this threshold. We already have a config
option for it. So let's reuse it.
check-config.py doesn't like having multiple defaults for the
ui.configwith() calls. The behavior is obviously correct. I'm
not sure if it is worth teaching check-config.py how to ignore
this. So I've just accepted the new output.
Context manager has a mechanism to control extension propagation. It is not
used by profiling right now, but making the code correct will help prevent bug
in the future.
There are no way for this to happen today, but better be safe than sorry, no
one know how the code will evolve. We now make sure the file pointer is closed
even is profiler is None.
It seems sufficiently simple to use "profile(enabled=X)" to not justify having
a dedicated context manager just to read the config.
(I do not have a too strong opinion about this).
This is a step toward allowing context where the profiling in enabled
withing the context range.
This also open the way to kill the dedicated "maybeprofile" context manager
and keep only one of 'profile' and 'maybeprofile'.
The start method is doing all profiler setup and activation. It is currently
unconditionally called by '__init__' but this will be made more flexible in
later changesets.
So far we have been able to use a simple decorator for this. However using the
current context manager makes the scope of the profiling in dispatch
constrainted and the time frame to decide to enable profiling quite limited
(using "maybeprofile")
This is the first step toward the ability to enable the profiling from within
the profiling scope. eg::
with maybeprofiling(ui) as profiler:
...
bar.foo():
...
if options['profile']:
profiler.start()
...
fooz()
...
My target usecase is adding support for "--profile" to alias definitions with
effect. These are to be used with "profiling.output=blackbox" to gather data
about operation that get slow from time to time (eg: pull being minutes instead
of seconds from time to time).
Of course, in such case, the scope of the profiling would be smaller since
profiler would be started after running extensions 'reposetup' (and other
potentially costly logic), but these are not relevant for my target usecase
(multiple second commits, multiple tens of seconds pull).
Currently adding '--profile' to a command through alias requires to re-spin a
Mercurial binary (using "!$HG" in alias), which as a significant performance
impact, especially in context where startup performance is being worked on...
An alternative approach would be to stop using the context manager in dispatch
and move back to a try/finally setup.
459366b580cf makes profiler start early without loading extensions. That
makes it impossible for an extension to add customized profilers.
This patch adds a special case: if a profiler is not found but an extension
with the same name could be loaded, load that extension first, and expect it
to have a "profile" contextmanager method. This allows customized profilers
and extension setup time is still profiled.
We synthesize function call begin/end events from snapshots, and
try (configurably) to eliminate "noisy" stack frames.
Example invocation:
hg --config profiling.output=$HOME/Desktop/clone.json \
--config profiling.statformat=chrome \
--profile clone https://www.mercurial-scm.org/repo/hg
pycompat.getenv returns os.getenvb on py3 which is not available on Windows.
This patch replaces them with encoding.environ.get and checks to ensure no
new instances of os.getenv or os.setenv are introduced.
os.getenv deals with unicodes on Python 3, so we have pycompat.osgetenv to
deal with bytes. This patch replaces occurrences on os.getenv with
pycompat.osgetenv
The statprof sampling profiler runs with significantly less overhead.
Its data is therefore more useful. Furthermore, its default output
shows the hotpath by default, which I've found to be way more useful
than the default profiler's function time table.
There is one behavioral regression with this change worth noting:
the statprof profiler currently doesn't profile individual hgweb
requests like lsprof does. This is because the current implementation
of statprof only profiles the thread that started profiling.
The ability for lsprof to profile individual hgweb requests is
relatively new and likely not widely used. Furthermore, I have plans
to modify statprof to support profiling multiple threads. I expect
that change to go through several iterations. I'm submitting this
patch first so there is more time to test statprof. Perfect is the
enemy of good.
Now that the statprof module is vendored and suitable for use, we
switch our statprof profiler to use it. This required some minor
changes because of drift between the official statprof profiler
and the vendored copy.
We also incorporate Facebook's improvements from the "statprofext"
extension at
https://bitbucket.org/facebook/hg-experimental, notably support for
different display formats.
Because statprof output is different, this is marked as BC. Although
most users likely won't notice since most users don't profile.
And refactor dispatch.py to use it. As you can see, the resulting code
is much simpler.
I was tempted to inline _runcommand as part of writing this series.
However, a number of extensions wrap _runcommand. So keeping it around
is necessary (extensions can't easily wrap runcommand because it calls
hooks before and after command execution).
This makes profiling more flexible since we can now call multiple
functions when a profiler is active. But the real reason for this
is to enable a future consumer to profile a function that returns
a generator. We can't do this from the profiling function itself
because functions can either be generators or have return values:
they can't be both. So therefore it isn't possible to have a generic
profiling function that can both consume and re-emit a generator
and return a value.
Currently, profiling code lives in dispatch.py, which is a low-level
module centered around command dispatch. Furthermore, dispatch.py
imports a lot of other modules, meaning that importing dispatch.py
to get at profiling functionality would often result in a module import
cycle.
Profiling is a generic activity. It shouldn't be limited to command
dispatch. This patch moves profiling code from dispatch.py to the
new profiling.py. The low-level "run a profiler against a function"
functions have been moved verbatim. The code for determining how to
invoke the profiler has been extracted to its own function.
I decided to create a new module rather than stick this code
elsewhere (such as util.py) because util.py is already quite large.
And, I foresee this file growing larger once Facebook's profiling
enhancements get added to it.
