Experimenting with tree inliner parameter settings

[Daniel Berlin]

On Sun, 4 May 2003, Jakub Jelinek wrote:

> On Sun, May 04, 2003 at 10:44:45PM +0200, Richard Guenther wrote:
> > [Summary / analysis of inline params snipped]
> >
> > The SPEC results show, that at least defaults for C and Fortran do not
> > necessarily apply to C++ code (as the testcases from the PRs show).
> >
> > So may I suggest again splitting the defaults per language? Especially for
> > templated C++ code with exceptions the instruction counting can be _very_
> > wrong.
>
> I strongly second this.
It's very hard to do with the current Tree IR, because the trees can
be of unbounded complexity.

For GIMPLE (in tree-ssa), because of the regularity of the IR, the
variation is much lower (it's 1-3 INSN's per stmt, rather than 1-15).

> Furthermore, I still think that e.g. in C explicit inline should be given
> bigger limits than -O3 inlining (and nested inline functions even bigger).
>
> 	Jakub
>