The cost of stack traces

[David Daney]

Andrew Haley wrote:
> David Daney writes:
>  > Andrew Haley wrote:
>  > > One thing that seems almost to have got completely lost in all this
>  > > talk about how to make stack traces faster is the sheer pointlessness
>  > > of it all.  Many (if not most) of the installed gcj-compiled libraries
>  > > don't have debuginfo and usually no-one cares about the line numbers
>  > > anyway.
>  > 
>  > I take exception with that last part :->
> 
> I stand by it.  Usually, gcj-compiled programs are *used*.

I know that.  That is why my priorities are sometimes different than 
yours.  My users are all software engineers that like stack traces that 
reveal as much information as possible.

> 
>  > Really the whole reason that stack traces exist is to make
>  > debugging easier (or even possible).  I am not trying to be
>  > pedantic about having libgcj behave identically to the JDK.  There
>  > are countless times (several per day) where by looking at a stack
>  > trace I have discovered the cause of some bug.  Without the stack
>  > trace I would have to run the program in gdb and hack around for a
>  > while.  It wouldn't surprise me if stack traces saved me an average
>  > of an hour a day.
> 
> That's beside the point.  I wasn't proposing to turn stack traces off.
> I was proposing to turn filenames/line numbers off, by default.
> 
>  > I would be almost as happy if instead of file/line number pairs, that 
>  > raw addresses were produced, but the new StackTraceElement 
>  > infrastructure has made that less easy (currently impossible I think).
> 
>  > > One thing we could do is detect libraries with no debuginfo and from
>  > > that point onwards not bother looking for it.
>  > > 
>  > > By all means let's make the traces faster but there are surely better
>  > > uses of our time.
>  > 
>  > We all have different priorities.  I know that WRT libgcj that
>  > yours do not exactly coincide with mine.
>  > 
>  > One of my priorities is to make it possible to fix bugs in large
>  > java programs running on Linux systems with less than 32MB of RAM
>  > and no swap devices.  Often we have to strip all debugging
>  > information, so having a line number decoder (based on addr2line or
>  > some built-in decoder) does not help.
> 
> So why do you care about line numbers?

As defined in this e-mail thread, I now realize that do don't care about 
'line-numbers'.  Thanks for helping me arrive at this conclusion.

What would you think about a patch that optionally fills in the fileName 
  property of StackTraceElement with a string representation of the IP 
and information obtained from dladdr related to that address?

That would make for easy and accurate off-line analysis of stack traces 
with very little runtime overhead.

David Daney.