Small documentation on how to get some meaningful results on where we spend our time:

Step 1: Building

Start from a fully optimized build and add the following flags

• -ggdb for debug symbols
• -fno-omit-frame-pointer to not have unidentifiable samples in the dataset

Step 2: Running

Start the program (for sake of simplicity: sequential program) through the following wrapper:

perf record --call-graph fp ./program

Step 3: Analyzing

For a live analysis try (in the same directory you ran stuff in):

perf report --hierarchy

Unfortunately, PDELab types are so long, that perf completely trips over and cuts them unreadable. To get around, do:

perf report --hierarchy --stdio --no-demangle > undemangled
c++filt < undemangled > demangled

And try to read the demangled file.

Step 4: Postprocessing

Identifying a stackframe between all the PDELab types can be really hard. One possibility I found:

cat undemangled | clang-format > formatted

Note, however that clang-format started sqapping on my laptop for Stokes types.