Interactive Tracing and Debugging of Calls to a Function or Method

Description

A call to trace allows you to insert debugging code (e.g., a call to browser or recover) at chosen places in any function. A call to untrace cancels the tracing. Specified methods can be traced the same way, without tracing all calls to the generic function. Trace code (tracer) can be any R expression. Tracing can be temporarily turned on or off globally by calling tracingState.

Usage

trace(what, tracer, exit, at, print, signature,
      where = topenv(parent.frame()), edit = FALSE)
untrace(what, signature = NULL, where = topenv(parent.frame()))

tracingState(on = NULL)
.doTrace(expr, msg)
returnValue(default = NULL)

Arguments

Details

The trace function operates by constructing a revised version of the function (or of the method, if signature is supplied), and assigning the new object back where the original was found. If only the what argument is given, a line of trace printing is produced for each call to the function (back compatible with the earlier version of trace).

The object constructed by trace is from a class that extends "function" and which contains the original, untraced version. A call to untrace re-assigns this version.

If the argument tracer or exit is the name of a function, the tracing expression will be a call to that function, with no arguments. This is the easiest and most common case, with the functions browser and recover the likeliest candidates; the former browses in the frame of the function being traced, and the latter allows browsing in any of the currently active calls. The arguments tracer and exit are evaluated to see whether they are functions, but only their names are used in the tracing expressions. The lookup is done again when the traced function executes, so it may not be tracer or exit that will be called while tracing.

The tracer or exit argument can also be an unevaluated expression (such as returned by a call to quote or substitute). This expression itself is inserted in the traced function, so it will typically involve arguments or local objects in the traced function. An expression of this form is useful if you only want to interact when certain conditions apply (and in this case you probably want to supply print = FALSE in the call to trace also).

When the at argument is supplied, it can be a vector of integers referring to the substeps of the body of the function (this only works if the body of the function is enclosed in { ...}). In this case tracer is not called on entry, but instead just before evaluating each of the steps listed in at. (Hint: you don't want to try to count the steps in the printed version of a function; instead, look at as.list(body(f)) to get the numbers associated with the steps in function f.)

The at argument can also be a list of integer vectors. In this case, each vector refers to a step nested within another step of the function. For example, at = list(c(3,4)) will call the tracer just before the fourth step of the third step of the function. See the example below.

Using setBreakpoint (from package utils) may be an alternative, calling trace(...., at, ...).

The exit argument is called during on.exit processing. In an on.exit expression, the experimental returnValue() function may be called to obtain the value about to be returned by the function. Calling this function in other circumstances will give undefined results.

An intrinsic limitation in the exit argument is that it won't work if the function itself uses on.exit with add= FALSE (the default), since the existing calls will override the one supplied by trace.

Tracing does not nest. Any call to trace replaces previously traced versions of that function or method (except for edited versions as discussed below), and untrace always restores an untraced version. (Allowing nested tracing has too many potentials for confusion and for accidentally leaving traced versions behind.)

When the edit argument is used repeatedly with no call to untrace on the same function or method in between, the previously edited version is retained. If you want to throw away all the previous tracing and then edit, call untrace before the next call to trace. Editing may be combined with automatic tracing; just supply the other arguments such as tracer, and the edit argument as well. The edit = TRUE argument uses the default editor (see edit).

Tracing primitive functions (builtins and specials) from the base package works, but only by a special mechanism and not very informatively. Tracing a primitive causes the primitive to be replaced by a function with argument ... (only). You can get a bit of information out, but not much. A warning message is issued when trace is used on a primitive.

The practice of saving the traced version of the function back where the function came from means that tracing carries over from one session to another, if the traced function is saved in the session image. (In the next session, untrace will remove the tracing.) On the other hand, functions that were in a package, not in the global environment, are not saved in the image, so tracing expires with the session for such functions.

Tracing an S4 method is basically just like tracing a function, with the exception that the traced version is stored by a call to setMethod rather than by direct assignment, and so is the untraced version after a call to untrace.

The version of trace described here is largely compatible with the version in S-Plus, although the two work by entirely different mechanisms. The S-Plus trace uses the session frame, with the result that tracing never carries over from one session to another (R does not have a session frame). Another relevant distinction has nothing directly to do with trace: The browser in S-Plus allows changes to be made to the frame being browsed, and the changes will persist after exiting the browser. The R browser allows changes, but they disappear when the browser exits. This may be relevant in that the S-Plus version allows you to experiment with code changes interactively, but the R version does not. (A future revision may include a ‘destructive’ browser for R.)

Value

In the simple version (just the first argument), trace returns an invisible NULL. Otherwise, the traced function(s) name(s). The relevant consequence is the assignment that takes place.

untrace returns the function name invisibly.

tracingState returns the current global tracing state, and possibly changes it.

When called during on.exit processing, returnValue returns the value about to be returned by the exiting function. Behaviour in other circumstances is undefined.

Note

Using trace() is conceptually a generalization of debug, implemented differently. Namely by calling browser via its tracer or exit argument.

The version of function tracing that includes any of the arguments except for the function name requires the methods package (because it uses special classes of objects to store and restore versions of the traced functions).

If methods dispatch is not currently on, trace will load the methods namespace, but will not put the methods package on the search list.

References

Becker, R. A., Chambers, J. M. and Wilks, A. R. (1988) The New S Language. Wadsworth & Brooks/Cole.

See Also

browser and recover, the likeliest tracing functions; also, quote and substitute for constructing general expressions.

Examples

require(stats)

##  Very simple use
trace(sum)
hist(rnorm(100)) # shows about 3-4 calls to sum()
untrace(sum)

## Show how pt() is called from inside power.t.test():
if(FALSE)
  trace(pt) ## would show ~20 calls, but we want to see more:
trace(pt, tracer = quote(cat(sprintf("tracing pt(*, ncp = %.15g)\n", ncp))),
      print = FALSE) # <- not showing typical extra
power.t.test(20, 1, power=0.8, sd=NULL)  ##--> showing the ncp root finding:
untrace(pt)

f <- function(x, y) {
    y <- pmax(y, 0.001)
    if (x > 0) x ^ y else stop("x must be positive")
}

## arrange to call the browser on entering and exiting
## function f
trace("f", quote(browser(skipCalls = 4)),
      exit = quote(browser(skipCalls = 4)))

## instead, conditionally assign some data, and then browse
## on exit, but only then.  Don't bother me otherwise

trace("f", quote(if(any(y < 0)) yOrig <- y),
      exit = quote(if(exists("yOrig")) browser(skipCalls = 4)),
      print = FALSE)

## Enter the browser just before stop() is called.  First, find
## the step numbers

untrace(f) # (as it has changed f's body !)
as.list(body(f))
as.list(body(f)[[3]]) # -> stop(..) is [[4]]

## Now call the browser there

trace("f", quote(browser(skipCalls = 4)), at = list(c(3,4)))
## Not run: 
f(-1,2) # --> enters browser just before stop(..)

## End(Not run)

## trace a utility function, with recover so we
## can browse in the calling functions as well.

trace("as.matrix", recover)

## turn off the tracing (that happened above)

untrace(c("f", "as.matrix"))

## Not run: 
## Useful to find how system2() is called in a higher-up function:
trace(base::system2, quote(print(ls.str())))

## End(Not run)

##-------- Tracing hidden functions : need 'where = *'
##
## 'where' can be a function whose environment is meant:
trace(quote(ar.yw.default), where = ar)
a <- ar(rnorm(100)) # "Tracing ..."
untrace(quote(ar.yw.default), where = ar)

## trace() more than one function simultaneously:
##         expression(E1, E2, ...)  here is equivalent to
##          c(quote(E1), quote(E2), quote(.*), ..)
trace(expression(ar.yw, ar.yw.default), where = ar)
a <- ar(rnorm(100)) # --> 2 x "Tracing ..."
# and turn it off:
untrace(expression(ar.yw, ar.yw.default), where = ar)


if(requireNamespace("nlme")) { # almost always, as it is recommended
  ## trace calls to the function lm() that come from
  ## the nlme package.
  trace("lm", where = asNamespace("nlme"))
        lm    (len ~ log(dose) * supp, ToothGrowth) -> fit1  # NOT traced
  nlme::lmList(len ~ log(dose) | supp, ToothGrowth) -> fit2  # traced
  untrace("lm", where = asNamespace("nlme"))
}