-flto making program slower?

[Tobias Burnus]

Uri Moszkowicz wrote:
> I'm having trouble with link time optimization in my application. It
> is a large application that uses only basic C++ (no exceptions, no
> templates, no STL, no floating point, etc). Like many applications,
> the source files are compiled separately into object files. Some of
> those are combined into shared libraries. The shared libraries are
> then statically linked

I have to admit that I do not understand the last two sentences: Are the 
object files (.o) in a shared library (.so, .dyn, .dll) or in a static 
library (.a)?

> I simply added "-flto" to the GCC command to create object files:
>      g++ -Wall -pipe -O3 -flto -fno-strict-aliasing -mtune=generic
> --no-exceptions -fPIC -c some.cc
>
> I then added "-flto" to the final link command but not the shared libraries:
>    g++ -o exec -Xlinker some1.o some2.o -static some1.a some2.a
> -Wl,--wrap,open -flto

You should also enable optimizations at linkage time (-O3 -mtune=generic 
etc.); otherwise, LTO might not do as much optimizations as you would 
like it to do.

> I ran a benchmark of tests and the resulting execution time is now
> about 7% higher than it was without "-flto" added. Any suggestions for
> how to improve this result or why it may have gotten slower?

Except for the issue mentioned above,  think the most likely problem is 
that too much inlining is done. Inlining is not always profitable, e.g. 
not for

for (....)
   {
     if (something very unlikely)
       some_func()
     do something ...
   }

Here, inlining "some_func" will fill the cache with rarely executed 
code. The compiler (developers) try to apply inlining carefully (e.g. 
growth limits) but it's all based on heuristics.

With LTO you enable optimizations between files and, hence, inlining. On 
average, it should be profitable. But it might also lead to inlining 
functions which shouldn't be inlined.


> If it helps, when I add "-fuse-linker-plugin" I get this error:
>    g++: error: -fuse-linker-plugin is not supported in this configuration

I think without -fuse-linker-plugin no functions can be inlined from 
static libraries (.a) but only from object files (.o).

Tobias