Feature #1906
Kernel#backtrace: Objectifying Kernel#caller
| Status: | Assigned | Start date: | 08/07/2009 | |
|---|---|---|---|---|
| Priority: | Normal | Due date: | ||
| Assignee: | % Done: | 0% |
||
| Category: | core | |||
| Target version: | 2.0.0 |
Description
Inspired by nobu's recent refactoring of Kernel#rand`, several conversations with Ruby implementors about Kernel#caller, and Rubinius' Backtrace class, I've put together a _rough_ demo of how we could "objectify" Kernel#caller. It's at http://tinyurl.com/m9fdrn [github.com], along with some initial specs.
Rationale
=========
`caller` has two principle uses:
* Allowing users to display the backtrace at a given point, e.g. `puts caller`.
* Introspection to determine the callpath that lead the current method.
The first use is reasonably achievable with `caller`, as long as you don't want to do any formatting of the output. The second is hard because it requires parsing lines of the `caller` Array with regular expressions, and knowing what the various permutations of output imply. It would be easier if we could inspect the call stack with a Ruby-ish API. Further, this would allow alternative implementations to provide this functionality without having to reverse-engineer the output of `caller`. As a result, backtraces would become more useful and code using them more portable.
Name
====
The advantage of calling this feature #backtrace is that it's consistent with the usage of the term by Thread and Exception. This, however, could also be construed as a disadvantage because although identically named the output would be materially different. I'm not sure of the best approach in this regard.
API
===
A Kernel method named, for sake of argument, 'backtrace' which returns a Backtrace object. It can be treated like an Array, in the same way `caller` is, because it's an Enumerable. It also has shortcuts for accessing the most recent entry on the stack. Each line in the backtrace is represented by a Backtrace::Line object which has #file, #line, and #name accessors which correspond to the filename, the line number, and the method name, respectively. For example:
backtrace.name # The name of the method which invoked the current one as a Symbol
backtrace.file(2) # The absolute filename of the 3rd entry in the backtrace
backtrace.each do |line| # Yields Line objects
puts line.method
end
backtrace.lines.select {|l| l.method == :foo} # #lines returns an Array of Line objects
Simple stuff.
Weaknesses
==========
Ideally, #name (not called #method because of the clash with Object#method) would return a Method object. One of the many advantages of this would be that we could combine backtraces with Method#parameters to display the signatures of each method. Unfortunately, I can't see a non-hackish way to create Method objects from the output of `caller`, because I don't know which object the method is bound to. but if this were possible it would be useful.
I'm currently throwing away some of the output of `caller` because I don't completely understand it. We'll need to decide whether this would be useful to expose via the API, and if so how.
I'd prefer to return a File object for #file, but the majority of Ruby APIs return pathnames instead, so I've went with convention.
So is there any interest in this type of thing? Is it worth exploring further?
Related issues
History
Updated by wycats (Yehuda Katz) almost 3 years ago
Ha. I mentioned that this would be desirable in a different thread. Put
me down for an enthusiastic +1.
Have you looked at Rubinius' objectified backtrace?
-- Yehuda
Run Paint Run Run wrote:
> Feature #1906: Kernel#backtrace: Objectifying Kernel#caller
> http://redmine.ruby-lang.org/issues/show/1906
>
> Author: Run Paint Run Run
> Status: Open, Priority: Normal
> Category: core
>
> Inspired by nobu's recent refactoring of Kernel#rand`, several conversations with Ruby implementors about Kernel#caller, and Rubinius' Backtrace class, I've put together a _rough_ demo of how we could "objectify" Kernel#caller. It's at http://tinyurl.com/m9fdrn [github.com], along with some initial specs.
>
> Rationale
> =========
>
> `caller` has two principle uses:
>
> * Allowing users to display the backtrace at a given point, e.g. `puts caller`.
> * Introspection to determine the callpath that lead the current method.
>
> The first use is reasonably achievable with `caller`, as long as you don't want to do any formatting of the output. The second is hard because it requires parsing lines of the `caller` Array with regular expressions, and knowing what the various permutations of output imply. It would be easier if we could inspect the call stack with a Ruby-ish API. Further, this would allow alternative implementations to provide this functionality without having to reverse-engineer the output of `caller`. As a result, backtraces would become more useful and code using them more portable.
>
> Name
> ====
>
> The advantage of calling this feature #backtrace is that it's consistent with the usage of the term by Thread and Exception. This, however, could also be construed as a disadvantage because although identically named the output would be materially different. I'm not sure of the best approach in this regard.
>
> API
> ===
>
> A Kernel method named, for sake of argument, 'backtrace' which returns a Backtrace object. It can be treated like an Array, in the same way `caller` is, because it's an Enumerable. It also has shortcuts for accessing the most recent entry on the stack. Each line in the backtrace is represented by a Backtrace::Line object which has #file, #line, and #name accessors which correspond to the filename, the line number, and the method name, respectively. For example:
>
> backtrace.name # The name of the method which invoked the current one as a Symbol
> backtrace.file(2) # The absolute filename of the 3rd entry in the backtrace
> backtrace.each do |line| # Yields Line objects
> puts line.method
> end
> backtrace.lines.select {|l| l.method == :foo} # #lines returns an Array of Line objects
>
> Simple stuff.
>
> Weaknesses
> ==========
>
> Ideally, #name (not called #method because of the clash with Object#method) would return a Method object. One of the many advantages of this would be that we could combine backtraces with Method#parameters to display the signatures of each method. Unfortunately, I can't see a non-hackish way to create Method objects from the output of `caller`, because I don't know which object the method is bound to. but if this were possible it would be useful.
>
> I'm currently throwing away some of the output of `caller` because I don't completely understand it. We'll need to decide whether this would be useful to expose via the API, and if so how.
>
> I'd prefer to return a File object for #file, but the majority of Ruby APIs return pathnames instead, so I've went with convention.
>
> So is there any interest in this type of thing? Is it worth exploring further?
>
>
> ----------------------------------------
> http://redmine.ruby-lang.org
>
>
Updated by znz (Kazuhiro NISHIYAMA) about 2 years ago
- Target version set to 2.0.0
Updated by naruse (Yui NARUSE) 11 months ago
- Status changed from Open to Assigned
- Assignee set to matz (Yukihiro Matsumoto)
Updated by ddebernardy (Denis de Bernardy) 10 months ago
Is this (or #3917, which also looks neat) anything that might make it into ruby 1.9.3? I was wondering how to get the calling file's name earlier today without resorting to caller() -- which yields an unnecessarily large string array.
As an aside, there's this sender gem written in C here, in the meanwhile, which implements something very similar to the suggested backtrace() function:
https://github.com/Asher-/sender