How does covr work anyway?


Since releasing covr I have gotten a couple of requests to explain how it works. This post is adapted from a vignette I wrote to try and explain that.

Other coverage tools

Prior to writing covr, there were a handful of coverage tools for R code. R-coverage by Karl Forner and testCoverage by Tom Taverner, Chris Campbell, Suchen Jin were the two I was most aware of.

R-coverage

R-coverage provided a very robust solution by modifying the R source code to instrument the code for each call. Unfortunately this requires you to patch the R source and getting the changes upstreamed into the base R distribution would likely be challenging.

Test Coverage

testCoverage uses an alternate parser of R-3.0 to instrument R code and record whether the code is run by tests. The package replaces symbols in the code to be tested with a unique identifier. This is then injected into a tracing function that will report each time the symbol is called. The first symbol at each level of the expression tree is traced, allowing the coverage of code branches to be checked. This is a complicated implementation I do not fully understand, which is one of the reasons I chose to write covr.

Covr

Covr takes an approach in-between the two previous tools, modifying the function definitions by parsing the abstract syntax tree and inserting trace statements. These modified definitions are then transparently replaced in-place using C. This allows us to correctly instrument every call and function in a package without having to resort to alternate parsing or changes to the R source.

Modifying the call tree

The core function in covr is R/trace_calls.R. This function was adapted from pryr::modify_lang. This recursive function modifies each of the leaves (atomic or name objects) of a R expression by applying a given function to them. For non-leaves we simply call modify_lang recursively in various ways. The logic behind modify_lang and similar functions to parse and modify R’s AST is explained in more detail at Walking the AST with recursive functions.

modify_lang <- function(x, f, ...) {
  recurse <- function(y) {
    # if (!is.null(names(y))) names(y) <- f2(names(y))
    lapply(y, modify_lang, f = f, ...)
  }

  if (is.atomic(x) || is.name(x)) {
    # Leaf
    f(x, ...)
  } else if (is.call(x)) {
    as.call(recurse(x))
  } else if (is.function(x)) {
    formals(x) <- modify_lang(formals(x), f, ...)
    body(x) <- modify_lang(body(x), f, ...)
    x
  } else if (is.pairlist(x)) {
    # Formal argument lists (when creating functions)
    as.pairlist(recurse(x))
  } else if (is.expression(x)) {
    # shouldn't occur inside tree, but might be useful top-level
    as.expression(recurse(x))
  } else if (is.list(x)) {
    # shouldn't occur inside tree, but might be useful top-level
    recurse(x)
  } else {
    stop("Unknown language class: ", paste(class(x), collapse = "/"),
      call. = FALSE)
  }
}

We can use this same framework to instead insert a trace statement before each call by replacing each call with a call to a counting function followed by the previous call. Braces ({) in R may seem like language syntax, but they are actually a Primitive function and you can call them like any other function.

identical({1+2;3+4}, `{`(1+2, 3+4))
​[1] TRUE

Remembering that braces always return the value of the last evaluated expression we can call a counting function followed by the previous function substituting as.call(recurse(x)) in our function above with.

`{`(count(), as.call(recurse(x)))

Source References

Now that we have a way to add a counting function to any call in the AST we need a way to determine where in the code source that function came from. Luckily for us R has a built-in method to provide this information in the form of source references. When option(keep.source = TRUE) (the default for interactive sessions), a reference to the source code for functions is stored along with the function definition. This reference is used to provide the original formatting and comments for the given function source. In particular each call in a function contains a srcref attribute, which can then be used as a key to count just that call.

The actual source for trace_calls is slightly more complicated because we want to initialize the counter for each call while we are walking the AST and there are a few non-calls we also want to count.

Replacing

After we have our modified function definition how do we re-define the function to use the updated definition, and ensure that all other functions which call the old function also use the new definition? You might try redefining the function directly.

f1 <- function() 1

f1 <- function() 2
f1() == 2
​[1] TRUE

While this does work for the simple case of calling the new function in the same environment, it fails if the another function calls a function in a different environment.

env <- new.env()
f1 <- function() 1
env$f2 <- function() f1() + 1

env$f1 <- function() 2

env$f2() == 3
​[1] FALSE

As modifying external environments and correctly restoring them can be tricky to get correct, we use the C function reassign_function, which is used in testthat::with_mock(). This function takes a function name, environment, old definition, new definition and copies the formals, body, attributes and environment from the old function to the new function. This allows you to do an in-place replacement of a given function with a new function and ensure that all references to the old function will use the new definition.

S4 classes

R’s S3 and RC object oriented classes simply define functions directly in the packages namespace, so they can be treated the same as any other function. S4 methods have a more complicated implementation where the function definitions are placed in an enclosing environment based on the generic method they implement. This makes getting the function definition more complicated.

replacements_S4 <- function(env) {
  generics <- getGenerics(env)

  unlist(recursive = FALSE,
    Map(generics@.Data, generics@package, USE.NAMES = FALSE,
      f = function(name, package) {
      what <- methodsPackageMetaName("T", paste(name, package, sep = ":"))

      table <- get(what, envir = env)

      lapply(ls(table, all.names = TRUE), replacement, env = table)
    })
  )
}

replacements_S4 first gets all the generic functions for the package environment. Then for each generic function if finds the mangled meta package name and gets the corresponding environment from the base environment. All of the functions within this environment are then traced.

Compiled code

Gcov

Test coverage of compiled code uses a completely different mechanism than that of R code. Fortunately we can take advantage of Gcov, the built-in coverage tool for gcc and compatible reports from clang versions 3.5 and greater.

Both of these compilers track execution coverage when given the -fprofile-arcs -ftest-coverage flags. In addition it is necessary to turn off compiler optimization -O0, otherwise the coverage output is difficult or impossible to interpret as multiple lines can be optimized into one, functions can be inlined ect.

Makevars

R passes flags defined in PKG_CFLAGS to the compiler, however it also has default flags including -02 (defined in $R_HOME/etc/Makeconf) which need to be overridden. Unfortunately it is not possible to override the default flags with environment variables (as the new flags are added to the left of the defaults rather than the right). However if Make variables are defined in ~/.R/Makevars they are used in place of the defaults.

Therefore we need to temporarily add -O0 -fprofile-arcs -ftest-coverage to the Makevars file, then restore the previous state after the coverage is run. The implementation of this is in R/makevars.R.

Subprocess

The last hurdle to getting compiled code coverage working properly is that the coverage output is only produced when the running process ends. Therefore you cannot run the tests and get the results in the same R process. In order to handle this situation we use a subprocess function. This function aims to replicate as much of the calling environment as possible, then calls the given expressions in that environment and returns any objects created back to the calling environment. In this way we can have a fully transparent subprocess that acts like running the same code in the current process.

Compilation

The package’s code is compiled by calling devtools::load_all() with recompile=TRUE so while it does take some time to recompile all the code with tracing enabled, you do not have to do a full R CMD build cycle.

Running Tests

R code defined in tests or inst/tests is run using testthat::source_dir(), which simply sources each of the R files in those directories. This makes covr compatible with any testing framework.

Output Formats

The output format returned by covr is a very simple one. It consists of a named character vector, where the names are colon delimited information from the source references. Namely the file, line and columns the traced call is from. The value is simply the number of times that given call was called. There is also an as.data.frame method to make subsetting by various features simple.

While covr traces coverage by expression, typically users expect coverage to be reported by line.

Coveralls.io

Coveralls is a web service to help you track your code coverage over time, and ensure that all your new code is fully covered.

The coveralls API is fairly simple assuming you are already building on a CI service like travis. You send a JSON file to coveralls with the service_job_id (environment variable TRAVIS_JOB_ID), the service name (travis-ci) and a list of source files; each with the file name, file content and coverage per line for that file.

Once this JSON file is sent to the coveralls.io service it automatically generates coverage reports for the current state and tracks changes in test coverage over time.

Conclusion

Covr aims to be a simple easy to understand implementation, hopefully this vignette helps to explain the rational of the package and explain how and why it works.


Comments