Commit 283ec1da0f previously pushed had some coding style bugs, which
unfortunately couldn't be fixed before pushing because the platform
doesn't allow it and the original contributor is not available lately.
Let's fix these.
Continue installing the list of files to install even if a failure
occured. In particular, I encountered a problem where script-fu was
failing to install because the installed binary was being executed (and
its memory probably mmap-ed). Hence the copy failed with:
> [Errno 26] Text file busy: '/my/prefix/lib64/gimp/2.99/plug-ins/script-fu/script-fu'
Yet trying to reinstall plug-ins while GIMP is being run, hence
script-fu is being executed, is a very common development trick (for me
at least). So instead, when such an error occurs, I simply save it and
output all the exceptions in the end (where these warnings are properly
visible), instead of failing everything.
Of course, ideally the subtarget should not even try to install
script-fu as it is unchanged. This is why I changed shutil.copy() by
shutil.copy2() to install metadata, which include last access timestamp.
Maybe this could be used to decide whether the installed file is already
the last one, thus not retry, though I am actually unsure if this is
enough (on the other hand, the ultimate check of bit-to-bit comparison
may obviously be too much/slow which is counter-productive), which is
why I am not doing this change right now.
Add a new mnemonic-clashes tool, which checks for mnemonic clashes
in menus. This tool can be invoked directly from the shell. It
takes an optional parameter which limits the search to a specific
type of menus, according to their xml tag (in particular, "menu"
for regular menus, and "popup" for popup menus).
We should not rely on the current working directory, as it looks like
meson sets it to be the source directory (even when actually in the
build dir) when running custom target (so it fails).
Instead meson explicitly sets MESON_SOURCE_ROOT and MESON_BUILD_ROOT
environment variables for us. We should use these. This should work in
every cases (whether calling from the source or build dir).
Also removing the special-casing for a _build/ directory. This updated
version is generic and won't assume that you used some generic naming or
direct subdir under the source (even if many people might use this path
and name; I, for one, certainly don't!).
See: https://mesonbuild.com/Run-targets.html
Add a "markers" page to the performance-log viewer, which lists
the event markers contained in the log, and allows navigating
between them.
Update docs accordingly.
In the perofmance-log viewer, add header-bar buttons to clear and
invert the selection, and allow inverting the selection by ctrl-
right-clicking on the sample-selection area.
Update the docs.
In sample-search predicates, remove the "exclusive" parameter of
the "function()" function, and replace it with optional "id" and
"state" parameters, which limit the match to the call-stacks of
matching threads, as per the "thread()" function.
Sort the backtrace thread-list by thread ID.
In the performance-log viewer's profile view, allow sorting the
call-graph tree-views by function name, in addition to the
inclusive/exclusive frequencies.
In the performance-log viewer's profile view, displasy in-line bar-
chart-like visualization of function and source-line sample
percentages, as part of the corresponding tree-view cells.
In performance-log-viewer.py, fix thread-state toggling in the
profile-view thread-filter popover, when not all threads are
included in the current selection.
Add a new performance-log-coalesce.py tool, which groups together
performance-log address-map entries belonging to the same function
into a single symbol, by filling-in missing base symbol addresses.
The addresses are grouped such that each set of addresses
corresponding to a symbol of the same name, in the same source
file, are given the same (unique, but arbitrary) base address.
See the previous commit for why this is necessary.
This should work fine in most cases, however, for logs produced on
Windows, it can over-coalesce addresses belonging to different C++
lambda-functions in the same source file, since they all seem to be
given the same _FUN symbol name.
Use the new tool as part of the pipeline in performance-log-viewer.
Add an annotated source view to the performance-log viewer's
profile view. When selecting the [Self] entry of a function's
profile, for which source information is available and whose source
is found locally, a new column opens, showing the function's
source, annotated with sample statistics. Header-bar buttons allow
navigation through the annotated lines, selection of all the
samples corresponding to a given line, and opening the text editor
at the current line.
... when selecting a function's samples
Since we now preserve the call-graph path across state changes,
there's no need to explictly set the path after selecting a
function's samples in the profile view.
... in backtraces
In the performance-log viewer's backtrace viewer, show a document
icon next to stack frames with source-location information, whose
source file is found locally. Clicking the icon opens the source
file in a text editor at the relevant line.
Two environment variables control this feature:
- PERFORMANCE_LOG_VIEWER_PATH is a list of colon-separated
directories in which to look for source files. If this
variable is undefined, the current directory is used.
- PERFORMANCE_LOG_VIEWER_EDITOR is the command to use to launch
the text editor, for editing a specific file at a specific
line. The special strings "{file}" and "{line}" are replaced
with the filename and line-number, respectively. If this
variable is undefined, "xdg-open {file}" is used.
Don't take infinite values into account when calculating the
vertical scale of sample graphs, and rather display infinite values
as dashed lines at the top of the graph.
Fix int-ratio variable formatting when the input is NaN, which can
happen when calculating the standard deviation, if all the values
are infinite.
Fix keyboard sample-range selection.
Deselect all samples when right-clicking a sample graph.
In the performance-log viewer, add an option to filter which
threads, and which states of each thread, are included in the
profile. By default, all threads in the RUNNING state are
included.
In the performance-log viewer, defer updates to the various UI
elements when the selection changes until they're actually shown.
This improves responsiveness when changing the selection.
Use the less-ambiguous, if just as clumsy,
"format-indent-more/less" icons, to stand for the
caller -> callee, and callee -> caller, directions, respectively.
Fix searching for samples by thread name, in particular, when there
are unnamed threads.
Use GtkMenuButton, instead of GtkButton, for the find-samples
header button.
Add mnemonics to the find-samples popover.
performance-log-viewer.py is a viewer for GIMP performance logs.
The viewer is made up of two parts: a sample-selection area at the
top, and an information area at the bottom.
The sample-selection area visualizes the sampled variables and
markers using a simultaneous set of plots, and displays the
currently selected samples. Samples can be selected directly
through the sample-selection area, or by other means, such as
searching for all samples satisfying a certain condition, or
containing a certain function.
The information area shows global information stored in the log, as
well as information specific to the currently selected samples,
including variable listing and statistics, full backtrace, and
profile/call-graph information.
Note that performance-log-viewer.py takes its input from STDIN,
like the rest of the performance-log tools, and is therefore
suitable for use as part of a pipeline. For standalone use, the
performance-log-viewer driver is also included, which takes the log
file as a command-line argument, and processes it through an
appropriate pipeline before feeding it to the viewer.
The "running" attribute (readable through
gimp_backtrace_is_thread_running(), and recorded in the performance
log) specifies if the thread was in a running or suspended state at
the time the backtrace was taken. It is accurate on Linux, but
only approximated on Windows.
Adapt the performance-log-expand.py tool to maintain this attribute
(and any future thread attributes we might add).
luzpaz