Milian Wolff [Sun, 26 Feb 2017 21:05:20 +0000 (22:05 +0100)]
Exclude unresolved functions from recursion detection.
Previously, when we encountered a stack with two consecutive frames
with missing debug information, we thought it would indicate recursion
and then collapsed all frames beneath. Now, we exclude frames with
missing debug information and don't collapse the stacks anymore.
Milian Wolff [Sun, 26 Feb 2017 20:50:35 +0000 (21:50 +0100)]
Remove now-dead code
Milian Wolff [Sun, 26 Feb 2017 20:25:15 +0000 (21:25 +0100)]
Improve the bottom-up conversion to top-down and caller/callee data
When we encounter broken backtraces, we may not go all the way up
to main. In such cases, a non-leaf frame may actually have a cost
higher than the sum of its children. In these cases, we also have
to hande the difference in cost just like for a normal leaf node.
The manual test I added does not cover this properly, as the
unwinding is too reliable and thus we do not run into the corner case.
In real-world profiles though, this does occur and adds some errors.
This is now fixed and the caller/callee and top-down view become
more reliable.
l10n daemon script [Sun, 26 Feb 2017 06:51:10 +0000 (07:51 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Sun, 26 Feb 2017 05:46:07 +0000 (06:46 +0100)]
GIT_SILENT made messages (after extraction)
Milian Wolff [Thu, 23 Feb 2017 09:27:09 +0000 (10:27 +0100)]
Enable alternating row colors for caller/callee view
Milian Wolff [Thu, 23 Feb 2017 09:25:59 +0000 (10:25 +0100)]
Move top-view properties to .ui file, enable alternating row colors
l10n daemon script [Thu, 23 Feb 2017 07:13:09 +0000 (08:13 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Thu, 23 Feb 2017 06:13:29 +0000 (07:13 +0100)]
GIT_SILENT made messages (after extraction)
Milian Wolff [Wed, 22 Feb 2017 12:38:38 +0000 (13:38 +0100)]
Fix action enabled state
- only enable Open when we are not showing the open page already
- only disable Open and Close when we are not currently parsing a file
Milian Wolff [Wed, 22 Feb 2017 12:36:25 +0000 (13:36 +0100)]
Fix compile with older KF5, as e.g. available on Ubuntu 16.04
Milian Wolff [Tue, 21 Feb 2017 08:23:13 +0000 (09:23 +0100)]
Enable wordwrap in summary labels
When we encounter long debuggee command lines, we used to put them
into one line. This increased the minimal width of the mainwindow,
which could even expand beyond the screen dimensions, making parts
of the window invisible.
By enabling wordwrap, we can break the command line after every
argument. Additionally, to cope with extremely long arguments,
we add zero-width spaces after every 50 word characters, to enable
word wrap at these places.
BUG: 376741
Milian Wolff [Tue, 21 Feb 2017 07:54:13 +0000 (08:54 +0100)]
Document how to contribute to heaptrack
l10n daemon script [Mon, 20 Feb 2017 06:39:42 +0000 (07:39 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Mon, 20 Feb 2017 05:45:04 +0000 (06:45 +0100)]
GIT_SILENT made messages (after extraction)
l10n daemon script [Sun, 19 Feb 2017 06:37:47 +0000 (07:37 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Sat, 18 Feb 2017 06:55:48 +0000 (07:55 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Sat, 18 Feb 2017 06:01:16 +0000 (07:01 +0100)]
GIT_SILENT made messages (after extraction)
l10n daemon script [Fri, 17 Feb 2017 06:27:36 +0000 (07:27 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Fri, 17 Feb 2017 05:43:37 +0000 (06:43 +0100)]
GIT_SILENT made messages (after extraction)
l10n daemon script [Thu, 16 Feb 2017 06:38:37 +0000 (07:38 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Thu, 16 Feb 2017 05:45:11 +0000 (06:45 +0100)]
GIT_SILENT made messages (after extraction)
l10n daemon script [Wed, 15 Feb 2017 07:00:01 +0000 (08:00 +0100)]
SVN_SILENT made messages (.desktop file) - always resolve ours
In case of conflict in i18n, keep the version of the branch "ours"
To resolve a particular conflict, "git checkout --ours path/to/file.desktop"
l10n daemon script [Wed, 15 Feb 2017 06:02:53 +0000 (07:02 +0100)]
GIT_SILENT made messages (after extraction)
Milian Wolff [Mon, 13 Feb 2017 23:41:43 +0000 (00:41 +0100)]
Don't build the flamegraph twice on initial load
We get a resize event followed by a show event, both of which
used to trigger a rebuild of the flame graph. Now, we only react
once to that.
Milian Wolff [Mon, 13 Feb 2017 23:31:08 +0000 (00:31 +0100)]
Ensure the brushes are properly initialized in a thread-safe way
Fixes a crash under some circumstances, when multiple graphs
are build in parallel - something that needs to be investigated
separately.
Milian Wolff [Mon, 13 Feb 2017 23:29:33 +0000 (00:29 +0100)]
Set context menu policy on flamegraph main widget
Unbreaks the context menu, since the actions are set on the
main widget, but the policy was set on the view.
Milian Wolff [Mon, 13 Feb 2017 23:25:11 +0000 (00:25 +0100)]
Use KStandardAction for flamegraph forward/back actions.
Milian Wolff [Mon, 13 Feb 2017 23:25:45 +0000 (00:25 +0100)]
Use i18n in favor of QObject::tr
Milian Wolff [Mon, 13 Feb 2017 23:16:43 +0000 (00:16 +0100)]
apply clang-format
Milian Wolff [Mon, 13 Feb 2017 23:16:02 +0000 (00:16 +0100)]
Add open/open new/close/quit actions to file menu
Milian Wolff [Mon, 13 Feb 2017 22:14:12 +0000 (23:14 +0100)]
Add .desktop and appdata files
Milian Wolff [Sun, 12 Feb 2017 12:31:06 +0000 (13:31 +0100)]
Pass explicit parent to QAction ctor, fixes build on Ubuntu Xenial.
Seems like the default of `= nullptr` has been added recently to Qt.
BUG: 376247
Milian Wolff [Mon, 30 Jan 2017 20:55:19 +0000 (21:55 +0100)]
Optimize aggregation of caller/callee data
Previously, we did that on the unmerged data which means we had
to traverse way more stacks. Now, we instead use the merged data
as input and thereby can get rid of ~50% of the hash allocations
alone used for the recursion guard. This is the #1 hotspot when
it comes to allocation numbers so far, so that boils down to easily
millions of allocations.
Milian Wolff [Sun, 29 Jan 2017 19:41:37 +0000 (20:41 +0100)]
Show whether data comes from a runtime attached profile in GUI
Milian Wolff [Sun, 29 Jan 2017 19:35:30 +0000 (20:35 +0100)]
remove empty line
Milian Wolff [Sun, 29 Jan 2017 19:32:29 +0000 (20:32 +0100)]
Add format_sources for use with KDevelop's reformat-sources action.
Note: this depends on clang-format and moreutils' sponge
Milian Wolff [Sun, 29 Jan 2017 19:31:06 +0000 (20:31 +0100)]
Cleanup: add Elf namespace with typedefs for ElfW types
This looks much nicer in the C++ code below, and plays better with
clang-format as well.
Milian Wolff [Sun, 29 Jan 2017 19:25:42 +0000 (20:25 +0100)]
Fix warning about undefined behavior when expanding macro
See e.g. https://reviews.llvm.org/D15866 for more information
Milian Wolff [Sun, 29 Jan 2017 19:24:29 +0000 (20:24 +0100)]
remove dead code
Milian Wolff [Sat, 28 Jan 2017 22:46:10 +0000 (23:46 +0100)]
Set word-wrap-column to 120, like in clang-format config
Milian Wolff [Sat, 28 Jan 2017 22:41:50 +0000 (23:41 +0100)]
Add clang-format configuration and reformat src/ and test/ sources
Milian Wolff [Sat, 28 Jan 2017 21:51:33 +0000 (22:51 +0100)]
Remove libunwind-specific defines from global config
We have libunwind_config.h for that purpose.
Milian Wolff [Sat, 28 Jan 2017 21:49:09 +0000 (22:49 +0100)]
Fix detection of unw_set_cache_size and adapt to upstream changes
unw_set_cache_size is actually a define so we can detect it directly
in C++ without the need for CMake to detect it for us. Using the
define also makes sure we can detect it once for all platforms, the
symbol is mangled by libunwind to support multiple architectures.
Milian Wolff [Sat, 28 Jan 2017 20:56:11 +0000 (21:56 +0100)]
Move code to setup caller/callee view into separate function
Milian Wolff [Sat, 28 Jan 2017 20:52:51 +0000 (21:52 +0100)]
Share code to setup the TreeModel views
Milian Wolff [Sat, 28 Jan 2017 20:41:59 +0000 (21:41 +0100)]
Decouple heaptrack version from file format version.
This allows us to set the version again to 1.0.0 in preparation of
the first release. The file format is set to version 1 as well.
We are still able to read the old data files from profile runs
with current versions of heaptrack.
Milian Wolff [Sun, 22 Jan 2017 11:59:08 +0000 (12:59 +0100)]
Never show raw addresses in output.
This information is pretty useless, imho. Always returning a shared
<unresolved function> string as function name instead of the
stringified adress speeds up the code significantly and reduces
memory as more stack branches can get merged.
Milian Wolff [Sun, 22 Jan 2017 11:55:45 +0000 (12:55 +0100)]
Remove empty lines
Milian Wolff [Sun, 22 Jan 2017 11:37:34 +0000 (12:37 +0100)]
Speed up GUI analysis significantly by adding a location index cache
Perf showed that quite a lot of time is consumed finding the interned
location for a given InstructionPointer. By adding a cache in front
of this lookup that uses the IpIndex as key, the parse step speeds
up significantly, as indicated by the following perf stat calls,
which got measured by adding an `exit(0)` after buildSizeHistogram:
Before:
Performance counter stats for 'heaptrack_gui ./heaptrack.kwrite.27247.gz' (5 runs):
1384.265399 task-clock (msec) # 1.019 CPUs utilized ( +- 0.91% )
898 context-switches # 0.649 K/sec ( +- 1.44% )
77 cpu-migrations # 0.055 K/sec ( +- 23.67% )
27,747 page-faults # 0.020 M/sec ( +- 0.41% )
4,287,486,004 cycles # 3.097 GHz ( +- 0.90% )
5,965,906,048 instructions # 1.39 insn per cycle ( +- 0.07% )
1,298,988,220 branches # 938.395 M/sec ( +- 0.03% )
17,667,331 branch-misses # 1.36% of all branches ( +- 0.04% )
1.
358443950 seconds time elapsed ( +- 1.02% )
After:
Performance counter stats for 'heaptrack_gui ./heaptrack.kwrite.27247.gz' (5 runs):
969.782341 task-clock (msec) # 1.023 CPUs utilized ( +- 1.27% )
844 context-switches # 0.870 K/sec ( +- 6.82% )
98 cpu-migrations # 0.101 K/sec ( +- 5.30% )
27,590 page-faults # 0.028 M/sec ( +- 0.46% )
3,000,707,304 cycles # 3.094 GHz ( +- 1.22% )
4,374,580,309 instructions # 1.46 insn per cycle ( +- 0.03% )
911,989,553 branches # 940.406 M/sec ( +- 0.03% )
14,788,489 branch-misses # 1.62% of all branches ( +- 0.17% )
0.
947570109 seconds time elapsed ( +- 1.34% )
Milian Wolff [Sun, 22 Jan 2017 00:01:08 +0000 (01:01 +0100)]
Fix CMake warning about minimum version when trying to find KF5
Disable even trying to build the GUI when we deal with a CMake older
than 2.8.12.
CMake Warning (dev) at /home/milian/projects/compiled/kf5/share/ECM/modules/ECMFindModuleHelpers.cmake:133 (message):
Your project should require at least CMake 2.8.12 to use FindKF5.cmake
Call Stack (most recent call first):
/home/milian/projects/compiled/kf5/share/ECM/find-modules/FindKF5.cmake:52 (ecm_find_package_version_check)
CMakeLists.txt:35 (find_package)
This warning is for project developers. Use -Wno-dev to suppress it.
Milian Wolff [Sat, 21 Jan 2017 23:57:17 +0000 (00:57 +0100)]
Update copyright header in two more places
Milian Wolff [Sat, 21 Jan 2017 23:42:23 +0000 (00:42 +0100)]
Elide text of flamegraph items
This prevents the flickering when hovering an item with a long text.
That lead the label to wordwrap, increasing the height of the label,
decreasing the height of the view, moving the scene, thus the mouse
points to a different view and we would repeat again.
Now we simply elide the text on the right of the label. This works,
but it's cutting of the text too soon, since it cannot cope with
the HTML and thinks all of these chars will get painted...
Milian Wolff [Sat, 21 Jan 2017 22:48:46 +0000 (23:48 +0100)]
Use new LocationData::Ptr typedef everywhere
Milian Wolff [Sat, 21 Jan 2017 22:46:24 +0000 (23:46 +0100)]
Add context menu to open file location to model views
This just goes through QDesktopServices and does not yet allow
the user to configure what editor he wants to use to open the file.
Milian Wolff [Sat, 21 Jan 2017 18:35:44 +0000 (19:35 +0100)]
Properly connect filters to corresponding view proxy model
Burkhard Lück [Sat, 21 Jan 2017 16:29:29 +0000 (17:29 +0100)]
move Messages.sh to follow code changes
Milian Wolff [Sat, 21 Jan 2017 12:58:21 +0000 (13:58 +0100)]
Make it possible to run heaptrack from the build folder
By setting the correct CMake properties, the libraries, shell runner
and executables land in the correct folder layout in the build
directory. This way, we can run heaptrack directly from the build
dir, i.e.:
cd heaptrack/build
make
./bin/heaptrack <your app>
./bin/heaptrack_gui <data file>
Milian Wolff [Sat, 21 Jan 2017 12:40:21 +0000 (13:40 +0100)]
Move libunwind CMake logic into FindLibunwind.cmake
This fixes a build regression when using a clean build folder.
The find script is copied from zbackup, cf.:
https://raw.githubusercontent.com/zbackup/zbackup/master/cmake/FindLibUnwind.cmake
It was slightly adapted to find the symbols required for heaptrack
and not depend on symbols we don't actually need here.
While at it, extract the configuration defines out into its own
libunwind_config.h.
Milian Wolff [Sat, 21 Jan 2017 11:49:57 +0000 (12:49 +0100)]
Collapse recursive frames by default in flame graph
When enabled, this option cleans up the flamegraph in face of
recursive function calls. Note that this only skips recursion when
a function is calling itself directly - if there is another function
inbetween, the frames won't get collapsed.
Milian Wolff [Sat, 21 Jan 2017 11:37:24 +0000 (12:37 +0100)]
Reorder caller/callee columns
- put location first
- then show peak memory consumption and leaked cost
- followed by number of allocations and temporary allocations
- finally show throughput of total bytes allocated
- sort by peak memory consumption
Milian Wolff [Sat, 21 Jan 2017 11:30:55 +0000 (12:30 +0100)]
Don't increment caller/callee cost multiple times for recursive stacks
When we encounter a call stack with a recursive function call, we must
not increment the caller/callee cost more than once per stack. This
fixes the cost calculation for such stacks, which used to produce
costs that where sometimes beyond the actual peak cost.
Milian Wolff [Sat, 21 Jan 2017 11:29:22 +0000 (12:29 +0100)]
Remove extra ';'
Fixes compiler warnings in pedantic mode.
Fabrizio Oddone [Fri, 20 Jan 2017 21:52:43 +0000 (13:52 -0800)]
Include cinttypes where required.
This change is required to fix compilation errors under DevToolset4 on
CentOS with gcc 5.2.1, cf.:
https://www.softwarecollections.org/en/scls/rhscl/devtoolset-4/
Milian Wolff [Wed, 11 Jan 2017 15:06:38 +0000 (16:06 +0100)]
Update copyright year
Milian Wolff [Wed, 11 Jan 2017 13:50:24 +0000 (14:50 +0100)]
Restructure source code directory layout
Everything that's not 3rdparty or tests is now moved into the src
subfolder. That one in turn contains the following folders:
- util: shared helper functionality
- track: core functionality of heaptrack, i.e. the stuff that
gets injected into the client code
- interpret: does the first-level aggregation and DWARF translations
on the raw instruction pointers returned by tracking code
- analyze: high-level analyzers, i.e.:
print: the ASCII dumper
gui: the Qt/KDE GUI
Milian Wolff [Sun, 8 Jan 2017 23:00:51 +0000 (00:00 +0100)]
Merge remote-tracking branch 'paulmenzel/improve-readme.md'
Milian Wolff [Sun, 8 Jan 2017 22:58:22 +0000 (23:58 +0100)]
Fix flamegraph tooltip for top-level item
Use the item's "function" directly, as that one is actually a
tooltip-like label already for the top-level item.
Milian Wolff [Sun, 8 Jan 2017 22:46:10 +0000 (23:46 +0100)]
Add screenshots to README.md file
Milian Wolff [Sun, 8 Jan 2017 22:33:04 +0000 (23:33 +0100)]
Use the internal `__libc_dlopen_mode` symbol instead of `dlmopen`
Both, `dlmopen` and `dlopen` are only available when the debugeee
application links against libdl. When that is not the case, one
could not inject heaptrack at runtime.
By instead using the libc.so internal `__libc_dlopen_mode` function,
we get rid of this limitation and we can also inject into applications
that only link against libc.so but not libdl.
This trick was found in the linux-inject project:
https://github.com/gaffe23/linux-inject
Milian Wolff [Sun, 8 Jan 2017 22:33:01 +0000 (23:33 +0100)]
Document that you need gdb for runtime-attaching.
Milian Wolff [Sun, 8 Jan 2017 22:10:28 +0000 (23:10 +0100)]
Make dependency on KChart optional
The heaptrack_gui is pretty useable without KChart. So for those
who don't have access to KChart, simply skip charting and only
display the table views of raw data. Still better than using
heaptrack_print.
Milian Wolff [Sun, 8 Jan 2017 21:26:38 +0000 (22:26 +0100)]
Add basic documentation on how to build heaptrack
Most notably, tell the world about the various dependencies that
heaptrack has, both the core data collector as well as the analyzer
GUI.
Milian Wolff [Tue, 15 Nov 2016 22:29:54 +0000 (23:29 +0100)]
Link against libunwind directly, we don't need to link against the backtrace library here.
Milian Wolff [Tue, 15 Nov 2016 22:26:53 +0000 (23:26 +0100)]
Fix use-after-free in usage of libbacktrace.
We passed the c_str of a temporary std::string to libbacktrace,
which stored that in internal data structures. Later on, that dangling
pointer was accessed leading to subtle bugs which for odd reasons
did not became apparent until recently. Running the interpreter
through valgrind easily showed the problem though.
This fixes the 'failed to read executable information' error messages
emitted by heaptrack, and also ensures the profile data attributes
costs to the correct modules.
CCMAIL: apol@kde.org
Paul Menzel [Thu, 1 Sep 2016 13:30:37 +0000 (15:30 +0200)]
README: Remove trailing whitespace
Paul Menzel [Thu, 1 Sep 2016 13:31:33 +0000 (15:31 +0200)]
README: Add newline to end of file
Text files should end with a newline. So add one.
Milian Wolff [Tue, 16 Aug 2016 08:08:48 +0000 (10:08 +0200)]
Extend error message.
If dwarf.h is not found, then either elfutils-dev or libdwarf-dev need to be installed
Milian Wolff [Mon, 20 Jun 2016 09:42:59 +0000 (11:42 +0200)]
Add pool-allocator API for heaptrack.
This is a header-only API that you can use to add support for
pool-allocators or other custom allocators to heaptrack. Simply
include the header and call the `heaptrack_report_*` macros in
your custom allocator. When you trace your code under heaptrack,
the allocations will then be tracked. If you don't run under
heaptrack, nothing will happen.
#include <heaptrack.h>
void* MyPool::allocate(size_t s)
{
void* ptr = /* allocate s */;
heaptrack_report_malloc(ptr, s);
return ptr;
}
void MyPool::free(void* ptr)
{
heaptrack_report_free(ptr);
/* actually free */
}
Milian Wolff [Thu, 16 Jun 2016 10:21:37 +0000 (12:21 +0200)]
Make libheaptrack.h C-compliant.
Milian Wolff [Mon, 20 Jun 2016 11:52:16 +0000 (13:52 +0200)]
Skip top rows without cost.
This can happen when diffing files.
Milian Wolff [Fri, 17 Jun 2016 16:47:35 +0000 (18:47 +0200)]
Remove rows without cost from caller/callee data.
This reduces clutter for diffs.
Milian Wolff [Fri, 17 Jun 2016 16:28:00 +0000 (18:28 +0200)]
Merge branch 'wip/diffing'
Milian Wolff [Fri, 17 Jun 2016 16:27:22 +0000 (18:27 +0200)]
Also disable flame graphs in diff mode.
In the future, we may want to implement differential flame graphs,
but for now this is too much work.
Milian Wolff [Fri, 17 Jun 2016 15:38:32 +0000 (17:38 +0200)]
Disable charts in diff mode.
This could be implemented eventually, but is a lot of work.
Milian Wolff [Fri, 17 Jun 2016 15:23:14 +0000 (17:23 +0200)]
Load diff base data in parallel to primary data file.
Slightly improves performance when diffing.
Milian Wolff [Fri, 17 Jun 2016 14:41:23 +0000 (16:41 +0200)]
Reword --diff CLI help text, add -d, for heaptrack_gui.
Milian Wolff [Fri, 17 Jun 2016 14:38:45 +0000 (16:38 +0200)]
Make it possible to easily diff between files from the heaptrack_gui.
Milian Wolff [Fri, 17 Jun 2016 11:56:36 +0000 (13:56 +0200)]
Add some more code comments for the diffing method.
Milian Wolff [Tue, 14 Jun 2016 21:45:36 +0000 (23:45 +0200)]
Convert more data types to signed integers.
We now also calculate the difference between time, peakRSS and
system info when diffing heaptrack data files.
Milian Wolff [Tue, 31 May 2016 22:06:12 +0000 (00:06 +0200)]
Add diff support between files to heaptrack_{print,gui}.
Both executables get a new `--diff <file>` argument. The file
that is passed gets parsed in addition to the primary data file.
The allocation data therein will then be subtracted from the data
found in the primary data file.
The algorithm is somewhat complicated, because we have to ignore
instruction pointer addresses, as these will vary between runs
due to address space randomization. Paired with (partial) lack of
debug symbols, we end up with non-trivial situation to map allocations
from one run to that of another.
Essentially, we need to find the correct allocation that shares the
same backtrace while ignoring instruction pointer addresses.
Milian Wolff [Tue, 14 Jun 2016 15:20:29 +0000 (17:20 +0200)]
Don't crash when switching the call-stack direction.
Milian Wolff [Mon, 13 Jun 2016 09:13:11 +0000 (11:13 +0200)]
Add basic README.md
Milian Wolff [Mon, 30 May 2016 19:27:23 +0000 (21:27 +0200)]
Refactor code to use signed data types.
This is required to handle diffing between files or between
individual timestamps. Since we use 64bit data types, we
can still handle a lot of allocations this way (a bit more than
9 exabytes).
Milian Wolff [Tue, 31 May 2016 22:06:02 +0000 (00:06 +0200)]
Add minus operator to AllocationData
Milian Wolff [Wed, 8 Jun 2016 16:27:34 +0000 (18:27 +0200)]
Make it possible to skip building of heaptrack_gui.
For embedded, we never want to build the gui, and having the
find scripts spit out their errors there is just noisy.
Now, the recommended way to build heaptrack for embedded Linux is:
cmake -DCMAKE_TOOLCHAIN_FILE=... -DHEAPTRACK_BUILD_GUI=OFF
Milian Wolff [Wed, 8 Jun 2016 15:16:27 +0000 (17:16 +0200)]
Refactor code to remove reliance on undefined behavior.
Previously, the code got away with a c-style reinterpret_cast
in a constexpr function, which is not allowed according to the
standard. I think I was just lucky that the compiler didn't saw
through that yet.
Also, the code assumed (and enforced) that all function pointers
are sizeof(void*), which worked, but was a bit clunky.
Now, we simply get rid of the list of addresses, and iterate
manually over all hooks. With C++17's apply, this could be implemented
even nicer with a tuple and std::apply. That doesn't exist yet, so
we manually call the functions.
Milian Wolff [Wed, 8 Jun 2016 13:08:16 +0000 (15:08 +0200)]
Slightly cleanup code of preload hooks.
We instantiate directly with the decltype of a function
pointer now, so thereis no need to construct the pointer
via a secondary type anymore.
Milian Wolff [Wed, 8 Jun 2016 12:48:59 +0000 (14:48 +0200)]
Increase portability to systems that don't have cfree available.
cfree should not be used, according to its man page, but if it's
being used then we need to overload it. On systems where it is not
available, simply skip this step to keep heaptrack compiling.
Milian Wolff [Sat, 4 Jun 2016 17:36:02 +0000 (19:36 +0200)]
Track temporary allocations by their AllocationInfo index.
There is no need to output a boolean '1' for temporary allocations,
as we have the same information when looking at the AllocationInfo
index that corresponds to a unique size and location. I.e. we used
to output
+ 12
- 12 1
+ 42
- 43
But the same information is held by
+ 12
- 12
+ 42
- 43