I realized that I forgot to mention another repository related to my last post: llvm-tools. As the name suggests, this repository contains some useful tools based on my llvm-analysis library. The most interesting tool for people who aren't me is ViewIRGraph, which makes it easy to visualize several interesting program graphs (anything supported by llvm-analysis). The help output gives a reasonable breakdown:

  ViewIRGraph - View different graphs for LLVM IR modules in a variety of formats

  Usage: ViewIRGraph [-o|--output FILE/DIR] (-t|--type TYPE) [-f|--format FORMAT] FILE
    Generate the specified graph TYPE for FILE

  Available options:
    -h,--help                Show this help text
    -o,--output FILE/DIR     The destination of a file output
    -t,--type TYPE           The graph requested.  One of Cfg, Cdg, Cg, Domtree, Postdomtree
    -f,--format FORMAT       The type of output to produce: Gtk, Xlib, Html, Canon, XDot, Eps, Jpeg, Pdf, Png, Ps, Ps2, Svg.  Default: Gtk

The input to the tool is any bitcode file or C/C++ source file (source files will be compiled with clang or clang++). You can specify the type of program graph you want to see with the -t option. I've found the most useful to be the CFG, CDG, and Postdominator tree. The call graph (Cg) is also often useful to get an idea of how a program is structured. For all of the graph types except for the call graph, the tool creates one graph per function.

The format option, -f, determines how the graph will be displayed. The Gtk and Xlib options create X11 windows to show each generated graph. These are suitable for small inputs with just a few functions. The Html option is probably the most useful. It requires that the -o option be set and creates one .html file for each function in the input. Each file embeds the graph for the function as an SVG using the OpenLayers library to provide panning and zooming support. The other output formats generate images in the format indicated by their names. These are a bit less useful because most of those formats do not have viewers capable of handling large graphs in a useful way.

I should note that this tool uses graphviz (through the aptly-named Haskell graphviz library). This is wonderful code re-use, but be aware that huge graphs can end up being laid out in less-than-useful ways. The tool does use reasonable clustering specifications where possible, which does tend to help.

As a side note, I've been using optparse-applicative for command line argument parsing lately, including for ViewIRGraph. I'll write up a post about it sometime, but I'll just say for now that it is the nicest argument parsing library I have used. It also automatically generated the delightful --help output above.