Templight 2.0 - Template Instantiation Profiler and Debugger
Templight is a Clang-based tool to profile the time and memory consumption of template instantiations and to perform interactive debugging sessions to gain introspection into the template instantiation process.
Disclaimer: Templight is still at a very preliminary state. We hope to get as much early adoption and testing as we can get, but be aware that we are still experimenting with the output formats and some of the behaviors.
Table of Contents
- Templight Profiler
- Templight Debugger
- Getting Started
- Getting and Compiling Templight
- Invoking Templight
- Using the Templight Profiler
- Default Output Location
- Using the Templight Debugger
- Using Blacklists
- Inspecting the profiles
The templight profiler is intended to be used as a drop-in substitute for the clang compiler (or one of its variants, like clang-cl) and it performs a full compilation, honoring all the clang compiler options, but records a trace (or history) of the template instantiations performed by the compiler for each translation unit.
The profiler will record a time-stamp when a template instantiation is entered and when it is exited, which is what forms the compilation-time profile of the template instantiations of a translation unit. The total memory consumption can also be recorded at these points, which will skew the time profiles, but provide a profile of the memory consumed by the compiler during the template instantiations.
The profiler is enabled by the templight option
-profiler, and it supports the following additional templight options:
-stdout- Output template instantiation traces to standard output (mainly for piping / redirecting purposes). Warning: you need to make sure the source files compile cleanly, otherwise, the output will be corrupted by warning or error messages.
-memory- Profile the memory usage during template instantiations.
-safe-mode- Output Templight traces without buffering, not to lose them at failure (note: this will distort the timing profiles due to file I/O latency).
-ignore-system- Ignore any template instantiation located in system-includes (-isystem), such as from the STL.
-output=<file>- Write Templight profiling traces to . By default, it outputs to "current_source.cpp.trace.pbf" or "current_source.cpp.memory.trace.pbf" (if
-blacklist=<file>- Specify a blacklist file that lists declaration contexts (e.g., namespaces) and identifiers (e.g.,
std::basic_string) as regular expressions to be filtered out of the trace (not appear in the profiler trace files). Every line of the blacklist file should contain either "context" or "identifier", followed by a single space character and then, a valid regular expression.
The templight interactive debugger is also a drop-in substitute for the clang compiler, honoring all its options, but it will interrupt the compilation with a console prompt (stdin). The operations of the templight debugger are modeled after the GDB debugger and essentially works the same, with most commands being the same, but instead of stepping through the execution of the code (in GDB), it steps through the instantiation of the templates.
In short, the templight debugger is to template meta-programming what GDB is to regular code.
The debugger is enabled by the templight option
-debugger, and it supports the following additional templight options:
-ignore-system- Ignore any template instantiation located in system-includes (-isystem), such as from the STL.
-blacklist=<file>- Specify a blacklist file that lists declaration contexts (e.g., namespaces) and identifiers (e.g.,
std::basic_string) as regular expressions to be ignored by the debugger. Every line of the blacklist file should contain either "context" or "identifier", followed by a single space character and then, a valid regular expression.
NOTE: Both the debugger and the profiler can be used together. Obviously, if compilation is constantly interrupted by the interactive debugger, the time traces recorded by the profiler will be meaningless. Nevertheless, running the profiler during debugging will have the benefit of recording the history of the template instantiations, for later review, and can also record, with reasonably accuracy, the memory consumption profiles.
Getting and Compiling Templight
Templight must be compiled from source, alongside the Clang source code.
Follow the instructions from LLVM/Clang to get a local copy of the Clang source code.
Clone the templight repository into the clang directories, as follows:
(from top-level folder) $ cd clang/tools $ mkdir templight $ git clone <link-to-clone-templight-github-repo> templight
- Add the
templightsubdirectory to CMake:
(from top-level folder) $ cd clang/tools $ echo "add_clang_subdirectory(templight)" >> CMakeLists.txt
- (Re-)Compile LLVM / Clang: (same as the corresponding step in LLVM/Clang instructions)
(from top-level folder) $ mkdir build $ cd build $ cmake -DLLVM_ENABLE_PROJECTS=clang ../llvm/ $ make
- If successful, there should be a
templight++executable in the build/bin folder.
Templight is designed to be a drop-in replacement for the clang compiler. This is because in order to correctly profile the compilation, it must run as the compiler would, honoring all compilation options specified by the user.
Because of this particular situation, the options for templight must be specially marked with
-Xtemplight to distinguish them from other Clang options. The general usage goes like this:
$ templight++ [[-Xtemplight [templight-option]]|[clang-option]] <inputs>
Or, for the MSVC-compatible version of templight (analoguous to clang-cl) use as follows:
$ templight-cl [[-Xtemplight [templight-option]]|[clang-option]] <inputs> (OR) > templight-cl.exe [[-Xtemplight [templight-option]]|[clang-option]] <inputs>
For example, if we have this simple command-line invocation of clang:
$ clang++ -Wall -c some_source.cpp -o some_source.o
The corresponding templight profiler invocation, with options
-ignore-system would look like this:
$ templight++ -Xtemplight -profiler -Xtemplight -memory -Xtemplight -ignore-system -Wall -c some_source.cpp -o some_source.o
Note that the order in which the templight-options appear is not important, and that clang options can be interleaved with templight-options. However, every single templight-option must be immediately preceeded with
As can be seen from that simple example, one can easily substitude the clang command (e.g.,
clang++) with a compound templight invocation (e.g.,
templight++ -Xtemplight -profiler -Xtemplight -memory) and thus, leave the remainder of the command-line intact. That is how templight can be used as a drop-in replacement for clang in any given project, build script or an IDE's build configuration.
If you use CMake and Clang with the following common trick:
$ export CC=/path/to/llvm/build/bin/clang $ export CXX=/path/to/llvm/build/bin/clang++ $ cmake <path-to-project-root>
Then, templight could be swapped in by the same trick:
$ export CC="/path/to/llvm/build/bin/templight -Xtemplight -profiler -Xtemplight -memory" $ export CXX="/path/to/llvm/build/bin/templight++ -Xtemplight -profiler -Xtemplight -memory" $ cmake <path-to-project-root>
But be warned that the cmake scripts will not fully recognize this compiler, and therefore, you might have to make small changes in the CMake files to be able to handle it. Up to now, experience with building complete cmake projects with templight has been very successful as cmake essentially recognizes templight as a "Clang" compiler and applies the appropriate configuration. There were, however, a few minor issues found that were easily circumvented on a case-by-case basis, and so, be warned that this might not work perfectly on the first try. If anyone is interested in creating some CMake modules to deal with templight specifically, please contact the maintainer, such modules would be more than welcomed. Also, any feedback on your experience working with templight on a large project would be much appreciated.
Using the Templight Profiler
The templight profiler is invoked by specifying the templight-option
-profiler. By default, if no other templight-option is specified, the templight profiler will output only time profiling data and will output to one file per input source file and will append the extension
.trace.pbf to the output filename (object file output).
For example, running the following:
$ templight++ -Xtemplight -profiler -c some_source.cpp
will produce a file called
some_source.o.trace.pbf in the same directory as
It is, in general, highly recommended to use the
-ignore-system option for the profiler (and debugger too) because instantiations from standard or third-party libraries can cause significant noise and dramatically increase the size of the trace files produced. If libraries like the STL or most of Boost libraries are used, it is likely that most of the traces produced relate to those libraries and not to your own. Therefore, it is recommended to use
-isystem when specifying the include directories for those libraries and then use the
-ignore-system templight-option when producing the traces.
Warning: Trace files produced by the profiler can be very large, especially in template-heavy code. So, use this tool with caution. It is not recommended to blindly generate templight trace files for all the source files of a particular project (e.g., using templight profiler in place of clang in a CMake build tree). You should first identify specific parts of your project that you suspect (or know) lead to template instantiation problems (e.g., compile-time performance issues, or actual bugs or failures), and then create a small source file that exercises those specific parts and produce a templight trace for that source file.
Default Output Location
The location and names of the output files for the templight traces needs some explanation, because it might not always be straight-forward. The main problem here is that templight is meant to be used as a drop-in replacement for the compiler (and it actually compiles the code, exactly as the vanilla Clang compiler does it). The problem is that you can invoke a compiler in many different ways. On typical small tests, you would simply invoke the compiler on a single source file and produce either an executable or an object file. In other case, you might invoke it with a list of source files to produce either an executable or a library. And for "real" projects, a build system will generally invoke the compiler for each source file, producing an object file each time, and later invoking the linker separately.
Templight must find a reasonable place to output its traces in all those cases, and often, specifying an output location with
-output does not really solve the problem (like this problem: multiple source files, one output location?). The only really straight-forward case is with the
-stdout option, where the traces are simply printed out to the standard output (presumably to be piped elsewhere).
One Source File, One Object File (without
-output option): Whenever doing a simple compilation where each source file is turned into a single object file (i.e., when using the
-c option), then the templight tracer will, by default, output the traces into files located where the output object files are put and with the same name, but with the addition of the templight extensions. In other words, for compilation instructions like
-c some/path/first_source.cpp some/other/path/second_source.cpp, you would get traces files:
some/other/path/second_source.o.trace.pbf. Whichever options you specify to change the location or name of the generated object file, templight traces will follow. In other words, the default trace filenames are derived from the final output object-file names.
One Source File, One Object File (with
-output option): Because the
-output option cannot deal with multiple files specified, templight will ignore this option in this case.
One Source File, (Some Output) File: If you do any other operation that is like a simple compilation (no linking), such as using the
-S option to general assembly listings or using the
-ast-dump option, then the situation is analoguous to when you use the
-c option (produce object-files). The templight trace file names are simply derived from whatever is the final output file name (e.g.,
some_source.s.trace.pbf). If you invoke templight in a way that does not involve any kind of an output file (such as syntax-only or ast-dump to stdout), then templight falls back to deriving the trace file names from the source file names (e.g.,
The overall behavior in the above three cases is designed such that when templight is invoked as part of a compilation driven by a build-system (e.g., cmake, make, etc.), which is often done out-of-source, it will output all its traces wherever the build-system stashes the object-files, which is a safe and clean place to put them (e.g., avoid pollution in source tree, avoid possible file-permission problems (read-only source tree), etc..). It is also easy, through most build-systems to create scripts that will move or copy those traces out of those locations (that might be temporary under some setups) and put them wherever is more convenient.
Many Source Files, (Some Output) File: If you invoke the compilation of several source files to be linked into a single executable or library, then templight will merge all the traces into a single output file, whose name is derived from the executable or library name. If no output name is specified, the compiler puts the executable into
a.out, and templight will put its traces into
a.trace.pbf. If you specify an output file via the
-output option, then the trace will be put into that file.
Using the Templight Debugger
The templight debugger is invoked by specifying the templight-option
-debugger. When the debugger is launched, it will interrupt the compilation with a console command prompt, just like GDB. The templight debugger is designed to work in a similar fashion as GDB, with many of the same familiar commands and behavior.
Here is a quick reference for the available commands:
break <template name> |
b <template name> |
rbreak <regex> |
This command can be used to create a break-point at the instantiation (or any related actions, such as argument deductions) of the given template class or function, or any regular expression that matches with the instantiations that occur. If the base name of the template is used, like
my_class, then the compilation will be interrupted at any instantiation in which that base name appears. If an specialization for the template is used, like
my_class<double>, then the compilation will be interrupted only when this specialization is encountered or instantiated.
delete <breakpoint index> |
d <breakpoint index>:
This command can be used to delete a break-point that was previously created with the
break command. Each break-point is assigned an index number upon creation, and this index number should be used to remove that break-point. Use the
info break (or
i b) command to print the list of existing break-points so that the break-point index number can be retrieved, if forgotten.
This command (in either forms or their short-hands) will resume compilation until the next break-point is encountered.
This command (in either forms or their short-hands) will resume compilation until the end, ignoring all break-points. In other words, this means "finish the compilation without debugging".
This command will step into the template instantiation, meaning that it will resume compilation until the next nested (or dependent) template is instantiated. For example, when instantiating a function template definition, function templates or class templates used inside the body of that function will be instantiated as part of it, and the
step command allows you to step into those nested instantiations.
This command will skip to the end of the current template instantiation, and thus, skipping all nested template instantiations. The debugger will interrupt only when leaving the current instantiation context, or if a break-point is encountered within the nested template instantiations.
This command can be used to print the current stack of nested template instantiations.
info <kind> |
This command can be used to query information about the state of the debugging session. The
<kind> specifies one of the following options:
f: Prints the information about the current template instantiation that the debugger is currently interrupted at.
b: Prints the list of all existing break-points.
s: This is equivalent to calling
lookup <identifier> |
l <identifier> |
rlookup <regex> |
This command can be used to perform a look-up operation for a given identifier or regular expression, within the context of the current template instantiation. This will print out the match(es) that the compiler would make to the given identifier. For example, if you are in the instantiation of a member function of an STL container and lookup an identifier like
value_type, it will point you to the location of that the declaration that the compiler will associate to that identifier in that context (presumably a nested typedef in the STL container class template).
typeof <identifier> |
t <identifier> |
rtypeof <regex> |
This command is similar to
lookup except that instead of showing you the location of the matching identifier or regular expression (e.g., showing you a nested typedef or a data member) it will show you its type. If the identifier matches a variable, data member, function, or compile-time constant value, this command will show you the type of that object. If the identifier matches a type, it will show you what the actual type (fully instantiated) is. For example, if it matches a typedef or an alias, then the underlying type that it aliases will be shown (e.g., say you are inside the instantiation of the
std::vector<double>::insert function and issue
typeof value_type, it should show you
double as a result).
eval <identifier> |
e <identifier> |
reval <regex> |
This command can be used to perform a look-up and evaluation operation for a given identifier or regular expression, within the context of the current template instantiation. This will print out the compile-time value(s) of the match(es) that the compiler would make to the given identifier. For example, if you are in the instantiation of a member function of an
numeric_limits and lookup an identifier like
digits, it will give you the compile-time value of that member.
whois <identifier> |
w <identifier> |
rwhois <regex> |
This command can be used to perform the equivalent of
typeof operations all at once for a given identifier or regular expression, within the context of the current template instantiation. This will print out the match(es) that the compiler would make to the given identifier and the message will contain the identifier that the look-up resolves to, its value (if any) and its canonical type, and will point to the location of the identifier. This command is mostly provided as a convenient all-in-one option for manual debugging using this debugger, and thus, provides more "human-readable" messages for various situations. The output of the other three fundamental commands (lookup, eval and typeof) are structured for providing all information in a way that is easy to parse by, for example, a graphical front-end for the debugger.
setmode verbose |
This command can be used to toggle the verbose or quiet modes for the print outs. For example, in quiet mode (default), the location of declarations printed out during lookup or typeof operations will only name the source file and give the line / column numbers, but under verbose mode, it will also print out the corresponding line of source code and an indicator of the point (column) of the declaration (similar to diagnostic message of the compiler).
Warning: The templight debugger is still in an experimental phase. You should expect that some of the behaviors mentioned in the above descriptions of the commands will not work as advertized, yet. If you observe any unusual or buggy behaviors, please notify the maintainer and provide an example source file with the sequence of commands that led to the observed behavior.
A blacklist file can be passed to templight (profiler or debugger) to filter entries such that they do not appear in the trace files or do not get involved in the step-through debugging. The blacklist files are simple text files where each line contains either
context <regex> or
identifier <regex> where
<regex> is some regular expression statement that is used to match to the entries. Comments in the blacklist files are preceeded with a
The "context" regular expressions will be matched against declaration contexts of the entry being tested. For example, with
context std, all elements of the std namespace will be filtered out (but note that some elements of std might refer to elements in other implementation-specific namespaces, such as
__gnu_cxx for libstdc++ from GNU / GCC). Context blacklist elements can also be used to filter out nested instantiations. For example, using
context std::basic_string would filter out the instantiation of the member functions and the nested templates of the
std::basic_string class template, but not the instantiation of the class itself. In other words, declaration contexts are not only namespaces, but could also be classes (for members) and functions (for local declarations).
The "identifier" regular expressions will be matched against the fully-scoped entry names themselves. For example, using
identifier std::basic_string would filter out any instantiation of the
std::basic_string class template.
Warning: Beware of lexical clashes, because the regular expressions could blacklist things that you don't expect. For instance, using
context std would filter out things in a
lastday namespace too! It is therefore recommended to make the regular expressions as narrow as possible and use them wisely. For example, we could solve the example problem with
context ^std(::|$) (match only context names starting with "std" and either ending there or being followed immediately by "::"). Also, an often practical and safer alternative to blacklists is to use the
-ignore-system option, which ignores all entries coming out of a system-include header (standard headers and anything in directories specified with
-isystem compiler option), which is not as fine-grained but is often more effective against complex libraries (e.g., STL, Boost, etc.) which can cause a lot of "noise" in the traces.
Note that a pretty well-established convention for template-heavy code is to place implementation details into either
detail namespace or in an anonymous namespace. Boost library implementers are particularly good with this. Therefore, it can be a good idea to filter out those things if you are not interested in seeing instantiations of such implementation details.
Here is an example blacklist file that uses some of the examples mentioned above:
# Filter out anything coming from the std namespace: context ^std(::|$) # Filter out things from libstdc++'s internal namespace: context __gnu_cxx # Filter out anonymous entries (unnamed namespaces or types): identifier anonymous # Filter out boost::something::or::nothing::detail namespace elements: context ^boost::.*detail
Inspecting the profiles
To begin to inspect the profiles, the starting point is probably to head over to the sister repository called templight-tools. There, you will find utilities to deal with the trace files produced by templight. In particular, you can use
templight-convert to produce alternative formats, such as graphviz and callgrind, such that traces can be visualized. It is particularly recommended that you try out the "callgrind" output format, as it will allow the traces to be loaded in KCacheGrind for visualization.
Any contribution or work towards applications to help inspect, analyse or visualize the profiles is more than welcomed!
The Templar application is one application that allows the user to open and inspect the traces produced by Templight.
The Metashell project is another application that provides inspection facilities for templight trace files. It is a whole different application altogether (not strictly just a trace inspector), but invokes templight under the hood to generate the template instantiation tree that it allows you to walk, along with the AST.
The original version of Templight was created by Zoltán Borók-Nagy, Zoltán Porkoláb and József Mihalicza: http://plc.inf.elte.hu/templight/