[Bug tree-optimization/42209] missed optimization leads to several times slower code

[rguenth at gcc dot gnu dot org]

------- Comment #1 from rguenth at gcc dot gnu dot org  2009-11-29 12:54 -------
GCC has to weight code-size and compile-time increase against performance
improvements when deciding on inlining functions.  For the call in the loop
GCC assumes it is more beneficial to do the inlining compared to the calls
outside of the loop.  With the calls outside of the loop the limits set
for overall unit or function growth are reached.  You can get diagnostics
about this when you declare the function inline and use -Winline.

If you are sure that always inlining a function is beneficial you can
declare it with __attribute__((always_inline)).  You can also force
all calls in a function to be inlined by declaring that function
with __attribute__((flatten)).

GCC 4.5 inlines all calls if switch_case is declared inline, 4.4 produces

$ g++-4.4 -O3 t.C -Winline
t.C: In function ‘void slow(unsigned int*, const unsigned int*)’:
t.C:36: warning: inlining failed in call to ‘word_t switch_case(op_t, word_t,
word_t, word_t) [with word_t = unsigned int]Â’: call is unlikely and code size
would grow
t.C:67: warning: called from here
t.C:36: warning: inlining failed in call to ‘word_t switch_case(op_t, word_t,
word_t, word_t) [with word_t = unsigned int]Â’: call is unlikely and code size
would grow
t.C:79: warning: called from here

As 4.5 works it's unlikely to be fixed unless some of the profile fixes
also apply to 4.4 - Honza?


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu dot
                   |                            |org
          Component|c++                         |tree-optimization
           Keywords|                            |missed-optimization
      Known to work|                            |4.5.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42209