Re: [Injectj-devel] addition to analyse dead code?

[Thomas G. <Tho...@fz...>]

Hi *,

hmm, I don't know. Dead code analysis is either very conservative (i.e., 
   the statically observable control flow is analysed and some very 
rough estimations are made)  or very costly  (in case you use abstract 
interpretation; it basically boils down to points-to analysis which is 
undecidable in general and can thus only yield approximations).


Thus i don't really see what we would gain here. We either end up having 
the same functionality as the java compiler (in fact we already have 
this kind of stuff: "before exit" requires a control flow analysis as 
well as the check whether a fragment (the parameter to before/after 
operations) is valid, i.e., does not produce dead code). Or we need to 
integrate a sophisticated points-to analysis (in fact we have this 
analysis readily available: Holger Bär has built it for his PhD) which 
runs in the order of hours to produce decent results for a 
small-to-medium sized system.

A general-purpose trafo does not really make sense for my point of view 
(see the aforementioned problems).

We could however add a heuristic function to the, say, methods and 
classes like "boolean containsDeadCodeBlocks(...)" or even 
"markDeadCodeBlocks" to put comments around code which is proven to be 
"dead". Question is what the advantages are (or: do the advantages 
outweigh the disadvantages?)

The disadvantages I see:
- we need to be able to analyse dead code not only on the basis of 
static control flow (scf) approximations (i.e., the way of the java 
compiler). If we would only do scf analysis we either end up with no 
"hits" (given our pre-requisite -- the Java code is compilable before 
transformation -- still holds). Or (again with scf analysis) we drop 
this pre-requisite (i.e., if there is statically observable dead code in 
the program, it would never have made it through the compiler thus it is 
not correct) which I personally do not like either.

- Analysing the "dynamic" control flow is *very* costly. We would need 
to restrict ourselves to heuristics which are easy (and fast) to compute 
and still accurate enough to be of any use. If the belowmentioned tool 
can do this -- fine! If not I'd say let's drop this by now.

Cheers,
Thomas


P.S.: Don't drink to many B52's :-)


Sven Luzar wrote:
> Hello together,
> 
>  
> 
> the following webside presents an analyse Tool to detect dead code an so on.
> 
>  
> 
> http://pmd.sourceforge.net/
> 
>  
> 
> I think it’s a nice addition for Inject/J to use a tool like this.
> 
> We can use it for our own code and / or we can
> 
> support this information by Inject/J.
> 
> My idea is a transformation like this:
> 
>  
> 
> “Detect all dead code lines and add a comment for this lines.” or
> 
> “Detect all dead code lines and remove this lines.”
> 
>  
> 
> What do you think about it?
> 
>  
> 
> Mit freundlichem Gruß
> 
>  
> 
> Sven Luzar
> 
> Wasserstraße 23
> 
> 48565 Steinfurt
> 
>  
> 
> Tel: +49 (2551) 862250
> 
> Fax: +49 (89) 2443 68746
> 
> Email: Sve...@we...
> 
>  
> 

-- 
Thomas Genssler (Tho...@fz...)
Tel./Fax:       +49-721-9654-602/-603

Forschungszentrum Informatik - Research Center for Information Technologies
Programmstrukturen - Program Structures (http://www.fzi.de/prost)
Haid-und-Neu-Strasse 10-14
76131 Karlsruhe GERMANY