Task #1694
closedAvoid inline functions to reduce code size bloat
Added by Junxiao Shi over 10 years ago. Updated over 6 years ago.
0%
Description
Move all inline functions to .cpp, except trivial get/set accessors and templates.
Guidelines and reasons: C++ Dos and Don'ts
Files
name-noinline.patch (22.1 KB) name-noinline.patch | Junxiao Shi, 06/23/2014 01:09 AM |
Updated by Alex Afanasyev over 10 years ago
The "guidelines" have nothing to do with "size bloat". These guidelines are for the chromium project and I see no reason to adopt them.
There is no point of changing the current code, except wasting time and making excessive number of compilation units, wasting more time on the library compilation. The code "bloat" is irrelevant to the objectives of the project. The binary size will be large, since we are using many things from STL and boost that produce a lot of code (e.g., boost asio).
Updated by Junxiao Shi over 10 years ago
- File name-noinline.patch name-noinline.patch added
Based on commit 72a11780c76ca32ca574aaf1cfa18a733385ba78, constructors and long non-template functions in name.hpp
and name-component.hpp
are moved to .cpp files and made non-inline.
(see name-noinline.patch
for the modifications)
code | libndn-cxx.a unstripped size | libndn-cxx.a stripped size |
---|---|---|
inline | 71384210 | 4216674 |
non-inline | 71450052 | 4192898 |
Binary size is reduced by 23KB when inline constructors and long functions are avoided in Name
and name::Component
.
This illustrates that unnecessary inline functions can cause code size bloat.
In response to the worries in note-1, compilation time is also measured.
The commands used in my test are:
./waf distclean
git checkout <tested-branch>
./waf configure
time ./waf -j1
code | compilation time |
---|---|
inline | real 2m20.312s, user 2m7.936s, sys 0m11.393s |
non-inline | real 2m18.405s, user 2m6.284s, sys 0m11.897s |
This shows that avoiding inline functions does not increase compilation time.
All experiments are done on a Ubuntu 12.04 64-bit virtual machine.
The disk of this virtual machine is located on a SAS hard drive.
Updated by Alex Afanasyev over 10 years ago
Insignificant change. No reason to waste time on doing any updates.
The compilation time wouldn't change for just one change. But the consequence of your push is total removal of all inline-only classes, which is extremely bad idea.
I'm not changing the fact the most of the library stay as header only. This includes name, name-component, most of management, and many other core classes.
Updated by Junxiao Shi over 10 years ago
Compilation time is less important than binary size.
There's also no proof than compilation time is increased significantly.
Larger binary => more runtime memory consumption => (for embedded NFD) less memory for ContentStore.
Updated by Davide Pesavento over 10 years ago
Alex Afanasyev wrote:
There is no point of changing the current code, except wasting time and making excessive number of compilation units, wasting more time on the library compilation.
Actually, the opposite is true. More code in header files means that the compiler must go through a lot of the same code several times (once for each translation unit that directly or indirectly includes that header file), therefore increasing total build time. Precompiled headers may alleviate this issue but still won't reduce it to zero.
Moreover, the compiler will generate tons of duplicate symbols, which then the linker will have to merge and discard, further increasing build time.
I'm not changing the fact the most of the library stay as header only.
This is not a valid point IMHO. Having a header-only library could be a goal (however, who decided that it's a goal? and why?). But a "mostly" header-only library doesn't make sense, either all or nothing.
But the consequence of your push is total removal of all inline-only classes, which is extremely bad idea.
Why? You keep saying this without stating a provable reason.
Updated by Junxiao Shi over 10 years ago
20140708 conference call decides we should stop adding new inline functions unless they are trivial getters/setters.
Fixing old code is low priority, and @Beichuan can decide when to do it.
@Lixia Zhang points out "who cares" is not a right attitude toward this problem, because code size bloat does cause problem in small devices such as home routers.
Updated by Alex Afanasyev over 10 years ago
Please give the exact problem. What is the "target size"?
The few bytes that can be saved by non-inlining could be nothing compared to the amount of space you can save by switching to use shared library.
Updated by Junxiao Shi over 9 years ago
note-6 decision was made one year ago, and it has been caused numerous disputes.
I suggest amending the rule as:
- In general, don't use
inline
. - Constructors and destructors SHOULD NOT be
inline
(except template). - The following MAY be
inline
:- trivial getter, ie.
return m_field;
wherem_field
is a TriviallyCopyable type - trivial setter, ie.
m_field = arg;
wherem_field
is a TriviallyCopyable type - forwarding function, ie. call another function with same arguments; note: arguments should be TriviallyCopyable or accepted by
const T&
- small function body: no more than 1 function call (not counting calls to trivial getter/setter; including overloaded operator), no more than 7 operators, no control statements (such as
if switch for while throw try
)
- trivial getter, ie.
- A virtual method can be
inline
if satisfying above conditions.
Updated by Junxiao Shi over 9 years ago
20150801 conference call did not approve note-8.
The only approved rule is still note-6, which Alex very reluctantly agreed to a year ago.
However, I will use the loosen rules in note-8 for future code reviews.
Updated by Alex Afanasyev almost 9 years ago
- Target version changed from v0.3 to v0.5
Updated by Davide Pesavento over 7 years ago
- Priority changed from Normal to Low
- Target version deleted (
v0.5)
Updated by Davide Pesavento over 6 years ago
- Status changed from New to Closed
I'm closing this as there's no actionable item here.