-finline-limit

[Ian Lance Taylor]

Maurizio Vitale <maurizio.vitale@polymath-solutions.com> writes:

> In a program that does a lot of template metaprogramming stuff, I
> have a function
> that I need to be inlined. The resulting code (when inlining happens)
> is very short
> (about 10-20 instructions), but I have no idea of the complexity of
> functions the
> compiler is exposed to.
> 
> Now the fun thing: documentation says that the default for -finline-
> limit is 600
> and that's the number of ops (in the RTL, I presume).
> What happens is that compiling without -finline-limit (gcc 4.1.1)
> doesn't inline
> my function, but compiling w/ almost any setting (including much
> below the default of 600)
> does inline succesfully.
> Specifically, I have to go down to -finlinel-imit=60 to fail to
> inline and -finline-limit=65
> does inline succesfully).
> 
> Now I kind of remember from mailing lists a couple of years ago (but
> couldn't find
> the threads) that the scale for something (possibly -finline-limit)
> was changed
> (and the factor of 10 and the number 600 kind of resonate).
> 
> Anybody to comment?
> It is possible that the docs are wrong and the default is now 60?

The docs are wrong.

What they should actually say is more complicated.  -finline-limit now
controls two different parameters.  -finline-limit=N sets both
parameters to N / 2.  The defaults for the parameters are different
(one defaults to 450, the other to 90).  So -finline-limit=900 sets
one to the default and increases the other, while -finline-limit=180
decreases the first and sets the other to the default.

Ian