Archive

Posts Tagged ‘source transformation’

Semantic pattern matching (Coccinelle)

January 8, 2009 No comments

I have just discovered Coccinelle a tool that claims to fill a remarkable narrow niche (providing semantic patch functionality; I have no idea how the name is pronounced) but appears to have a lot of other uses. The functionality required of a semantic patch is the ability to write source code patterns and a set of transformation rules that convert the input source into the desired output. What is so interesting about Coccinelle is its pattern matching ability and the ability to output what appears to be unpreprocessed source (it has to be told the usual compile time stuff about include directory paths and macros defined via the command line; it would be unfair of me to complain that it needs to build a symbol table).

Creating a pattern requires defining identifiers to have various properties (eg, an expression in the following example) followed by various snippets of code that specify the pattern to match (in the following <… …> represents a bracketed (in the C compound statement sense) don’t care sequence of code and the lines starting with +/- have the usual patch meaning (ie, add/delete line)). The tool builds an abstract syntax tree so urb is treated as a complete expression that needs to be mapped over to the added line).

@@
expression lock, flags;
expression urb;
@@
 
  spin_lock_irqsave(lock, flags);
  &lt;...
- usb_submit_urb(urb)
+ usb_submit_urb(urb, GFP_ATOMIC)
  ...&gt;
  spin_unlock_irqrestore(lock, flags);

Coccinelle comes with a bunch of predefined equivalence relations (they are called isomophisms) so that constructs such as if (x), if (x != NULL) and if (NULL != x) are known to be equivalent, which reduces the combinatorial explosion that often occurs when writing patterns that can handle real-world code.

It is written in OCaml (I guess there had to be some fly in the ointment) and so presumably borrows a lot from CIL, perhaps in this case a version number of 0.1.3 is not as bad as it might sound.

My main interest is in counting occurrences of various kinds of patterns in source code. A short-term hack is to map the sought-for pattern to some unique character sequence and pipe the output through grep and wc. There does not seem to be any option to output a count of the matched patterns … yet 🙂