summaryrefslogtreecommitdiff
path: root/unusedcode.README
blob: c721d6237327f8fe935e38da361f3a2b641c8823 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
unusedcode.easy is generated via callcatcher[1] and filtered to remove some
tricky edge-cases (see Makefile), e.g. anything which could plausibly be
dlsymed or any symbol defined in an external library bundled into LibreOffice
which doesn't happen to get used by LibreOffice.

unusedcode.easy is generated on an x86_64 --enable-debug --enable-dbgutil
configuration.

Code listed as unused is code that gcc outputs but that nothing calls
(or takes the address of).

a) It's possible that some other platform or configuration uses the code,
   so manual inspection is always required.
b) At the time of writing the majority of unused code now originates via 
   macros, mostly from pre-STL containers, see [2] for killing two birds
   with one stone.
c) callcatcher ignores virtuals. But implementations of "pure virtuals"
   are not actually virtual methods. If something is declared pure virtual
   and provides an impl and that base-class impl is not explicitly called
   anywhere, then that impl can go away.
d) gcc will only emit code for inlines if those inlines are used, so
   sometimes something is listed correctly as unused but some inline
   code takes a pointer or reference to something which cannot be 
   instantiated so removal of some method/class fails at build time because
   gcc never emits any code for the the unused inline but trips over it at
   compile time after an attempt at removal. i.e. generally the inline method
   can go as well because nothing calls it either, a double win.
e) if a constructor is listed as unused, and it's the *only* ctor in the class,
   then no object of that class can be construsted, so the whole thing is
   unused, which can lead to a whole cascade of tricky but logical fallout.
f) if a destructor is listed as unused, and a constructor isn't, then there's
   a leak somewhere, and the destructor most likely *should* be called.
g) there's more actually unused code then what's listed. The idea is that what's
   listed is definitely unused under the generation configuration, not that
   it's a list of all unused code. If the count of unused easy hits 0 then
   we can have a look at the non-easy and if that hits 0, then strip out
   code from the "release" binaries which only makes sense in debug/dbgutil
   configurations, and then tackle unused virtual method slots :-)

[1] http://www.skynet.ie/~caolan/Packages/callcatcher.html
[2] https://bugs.freedesktop.org/show_bug.cgi?id=38832