00debug-package: How to use the debug package

Description Methods Special functions Options Display bugs Emacs Author(s) See Also


Update: as of v1.2.70, I think I may finally have a general-purpose fix for the tcltk bug that has caused display woes for many debug users. See Display bugs below.

debug is an alternative to trace and browser, offering:

Even if you don't write functions, or even if you don't write buggy functions, you may find it helpful to run the debugger on functions in package:base or other packages. Watching what actually happens while a function is executing, can be much more informative than staring at a piece of code or terse documentation.

Debugging your function f is a two-stage process. First, call mtrace(f) to store debugging information on f, and to overwrite f with a debug-ready version that will call the debugger itself. Second, do whatever you normally do at the command prompt to invoke f. This is often a direct call to f, but can be any command that eventually results in f being invoked. [The third, fourth, etc. stages, in which you actually fix the problem, are not covered here!]

When f is invoked, a window will appear at the bottom of the screen, showing the code of f with a highlight on the first numbered line. (There is also an asterisk in the far left hand column of the same row, showing that there's a breakpoint.) The command prompt in R will change to "D(...)> ", showing that you are "inside" a function that is being debugged. The debugger is now in "step mode". Anything you type will be evaluated in the frame of the function– this includes assignments and creation of new variables. If you just type <ENTER>, the highlighted statement in f will be executed. The result of the statement will be printed in the R command window, and the highlight will (probably) move in the f code window.

To progress through the code of f, you can keep pressing <ENTER>, but you can also type go() to put the debugger into "go mode", whereupon it will keep executing code statements without manual intervention. In "go mode", nothing will be printed in the command window (except if there are cat or print calls in the function code) until either:

In the first case, control is returned to the normal R command prompt, just as if the debugger had not been used. In the other cases, the D(...)> prompt will return and the line associated with the error / interrupt / breakpoint will be highlighted in the code window. You are then back in step mode. If there was an error, you can type statement(s) that will cause the error not to happen when the highlighted line executes again, or you can move the highlighted execution point to another line number by calling skip. Execution carries on quite normally after errors, just as if the offending statement had been wrapped in a try call. If your function eventually exits normally (i.e. not via qqq(), as described next), it will be as if the error never happened (though the error message(s) will be displayed when the R command prompt returns).

When in step mode, you can finish debugging and return to the normal R command prompt by typing qqq(). If you type <ESC> while in go mode, you should be returned to step mode, but sometimes you may be dumped back at the R command prompt (this is on to-be-fixed list), and sometimes there will be no immediate effect (e.g. if C code is running).

Breakpoints, including conditional breakpoints, are set and cleared by bp. Calling go(n) puts the debugger into go mode, but also sets a temporary breakpoint at line n, which will be triggered the first time execution reaches line n but not subsequently.

When the main function code has completed, the debugger moves into any on.exit code, which is also displayed and line-numbered in the code window. (Even if there are no calls to on.exit, a numbered NULL statement is placed in the exit code section, so that you can always set a "run-until-return" breakpoint.) If you exit via qqq(), the exit code will not be executed first; this can lead to subsequent trouble with open connections, screwed-up par values in graphics, etc.. To make sure the exit code does get executed:

When you want to restore f to its normal non-debugging state (and you are back at the real R command prompt), type mtrace(f,FALSE). To restore all debuggees, type mtrace.off(). It is advisable not to save functions in an mtraced state; to avoid manual untracing and retracing, look up Save in package mvbutils.

You can debug several functions "at once" (e.g. if f calls g, you can mtrace both f and g, with mtrace(g) called either inside or outside the debugger), causing several code windows to be open simultaneously. If f is called again inside f (either via some statement in f, or from something you type in step mode), another f code window will open. The number in the window title is the frame number, and the currently-active frame number is shown in the D(...)> prompt.

For statements typed in at the D(...)> prompt, only the first syntactically-complete R expression will be executed; thus, typing a <- 1; a <- 2 will set a to 1, but typing { a <- 1; a <- 2} will set a to 2.

See section SPECIAL.FUNCTIONS for handling of try, with, etc.

See section METHODS for how to handle S3 methods (easy), reference class methods (pretty easy), and S4 methods (not easy).

For further information, see R-news 3/3.


S3 methods work fine with mtrace; just do e.g. mtrace( print.classofthing). Reference class methods aren't too bad either— see ?mtrace for details. Unsurprisingly, S4 methods are much more painful to work with. I've only done it once; the following approach worked in R 2.12, but probably isn't optimal. Suppose you have figured out that the offending call is something like scrunge( x, y, z), where scrunge is the name of an S4 generic; e.g. you may have tried mtrace( scrunge), and found yourself with a debug window containing a 1-line function standardGeneric("scrunge"). First, use findFunction( "scrunge") to work out which package contains the definition of scrunge– say it's in package scrungepack. Next, you need to work out which specific scrunge method will be dispatched by scrunge( x, y, z). Try this:

  selectMethod( "scrunge", signature=character())
  # Look for the 'attr(,"target")' line; it might be e.g.
  # attr(,"target")
  #   x    y
  #  "ANY" "ANY"

Now you know that it's the x and y args that will direct traffic (it could have been just x, or just z, or...). So do class(x) and class(y) to figure out what the classes are; say they are matrix and character. Now do

  selectMethod( "scrunge", sig=c( "matrix", "character"))

Hopefully you'll see another attr(,"target") line, which will tell you which method to mtrace. Suppose it's the same as before (ANY and ANY); then you need to mtrace the ANY#ANY method. (If only one argument was used for dispatching, there wouldn't be a hash symbol.) The following magic formula will do it:

  mtrace( 'ANY#ANY', from=environment( scrungepack:::scrunge)$.AllMTable)

Then you can proceed as usual. Note that the method that is first dispatched may end up dispatching other methods– you will have to work out which yourself, and mtrace them accordingly. You can mtrace various functions in the methods package, e.g. callGeneric, which might help you track down what's going on. In short: good luck!

Special functions

Certain system functions with "code" arguments are handled specially in step-mode: currently try, suppressWarnings, eval, evalq, with, and within. In step-mode only, your code argument in these is stepped-into by default if it is more than a simple statement, using a new code window. In go-mode, and in step-mode if the default behaviour has been changed, these functions are handled entirely by R. Hence, if you are in go-mode and an error occurs in one of these statements, the debugger will stop at the with etc. statement, not inside it; but you can then step inside by pressing <ENTER>. The step-into-in-step-mode behaviour can be controlled globally using step.into.sysfuns. To avoid step-in at a specific line, you can also just use go to proceed to the following statement; this can be much faster than first stepping-in and then calling go, because R itself handles the evaluation.

To mimic the true behaviour of try, the debugger should really just return a "try-error" object if there is an error. However, that is probably not what you want when debugging. Instead, the debugger just behaves as usual with errors, i.e. it breaks into step-mode. If you do then want try to return the try-error, just as it would if you weren't debugging, type return( last.try.error()).

Note that the code window for with, eval, etc. contains an on-exit block, just like normal debugger code windows. Its main use is probably to let you set a breakpoint-on-return. However, it seems that you can successfully put on.exit statements inside your eval etc. call, whether debugging or not.

with and within are S3 generics, and the debug package only knows how to deal with the standard methods: currently with.default, within.data.frame, and within.list. You can debug specific methods manually, e.g. via mtrace( with.myS3class).

within is more complicated than the others, and two extra code windows are currently used: one for the code of within itself, and one for your statement. The operation of within itself is not usually very interesting, so the debugger does not pause except at the end of it, unless there is an error. Errors can occur during the updating of your object, which happens after your expression has been evaluated, e.g. from

  within( data.frame(), bad <- quote( symbols.not.allowed))


As of version 1.2.0, output is sent by default to stderr(); this means that you get to see the step-mode output even if the debuggee is redirecting "real" output (e.g. via sink). If you don't like this (and I believe Tinn-R doesn't), set options( debug.catfile="stdout").

Command recall is ON by default, but this means that anything typed while debugging will also be seen in history() after leaving the debugger. If this is a problem, set options( debug.command.recall=FALSE).

There are two adjustable limits on what gets printed out in step mode (otherwise, your screen will often fill with junk, and displaying results may take several minutes). First, printing will be aborted if it takes longer than getOption( "debug.print.time.limit") seconds, which by default is 0.5 seconds. You might need to increase that, e.g. for displaying big help files in the browser. Also, setting a finite time limit cutoff overrides any other time limits that have been set with setTimeLimit; this can be prevented by setting options( debug.print.time.limit=Inf). Second, by default only objects with object.size < 8192 bytes will be printed in full; for larger objects, a summary is given instead. You can force printing of any individual object via print, but you can also increase (or decrease) the threshold to X bytes, by setting options( threshold.debug.autoprint.size=X). The object.size test isn't foolproof, as some minuscule objects occupy hectares of screen real estate and take ages to print, whereas some big objects print compactly and fast. In my own work, I set the "threshold.debug.autoprint.size" option to Inf and the time limit usually to 0.5 seconds.

Various TCL/TK-related aspects of the code window can be altered:

If option debug.wordstar.keys is TRUE, various somewhat Wordstar-like key-bindings are provided: CTRL-E and CTRL-X to move blue selection up and down, CTRL-D and CTRL-S to scroll right/left, CTRL-W and CTRL-Z to scroll up/down, CTRL-R and CTRL-C to scroll up/down one page, and CTRL-K C to copy the current blue line to the clipboard (since CTRL-C has acquired a different meaning). Now that I've figured out how to add key bindings (sort of), more might appear in future.

Display bugs

Update: in v1.2.70, try setting options( shakeup.debug.windows=TRUE) if you have tcltk display woes. Please let me know if this doesn't solve it for you.

There have been sporadic and unreproduceable display problems with the TCL/TK window. Sometimes the window frame will appear, but with garbled contents or no contents at all. With RTERM in MS-Windows, a couple of ALT-TABs and mouse clicks to change focus are sometimes necessary. In extremis, the window will eventually sort itself out if you manually minimize, maximize, and restore it– admittedly an irritation. Although the problem seems less frequent these days, I'm not convinced it is fully solved.

Note: I was getting this problem all the time with R 2.13, after not having seen it at all since about R 2.8. It now seems OK again after making some random changes to the code; whatever the real problem is, it's probably not permanently fixed.

If you encounter this problem, try setting options(shakeup.debug.windows=TRUE) (also in .First). The price paid for setting this option, is that window focus will switch away from R to the new code window when the latter is first created, which is mildly irritating. On MS-Windows, I have a fix for that too, but it requires a DLL written in Delphi not C, so I can't CRAN it; let me know if you are desperate.

Apparently the RGtk2 package doesn't play nicely with debug– the debugging window displays a blank. I haven't checked this out, but would be grateful for advice.


For ESS users: I'm not an Emacs user and so haven't tried ESS with the debug package myself, but a read-through of the ESS documentation suggests that at least one ESS variable may need changing to get the two working optimally, as shown below. Please check the ESS documentation for further details on these points (and see also ?mvbutils). I will update this helpfile when I have more info on what works.


Mark Bravington

See Also

mtrace, go, skip, qqq, bp, get.retval, mtrace.off, check.for.tracees, step.into.sysfuns, last.try.error

debug documentation built on May 30, 2017, 2:23 a.m.

Related to 00debug-package in debug...