> On Wed, Apr 29, 2015 at 3:23 PM, Jan Hubicka <hubi...@ucw.cz> wrote:
> >> Historically the pragma Inline_Always of GNAT had been implemented in the
> >> FE
> >> because the RTL inliner and then the Tree inliner weren't invoked at -O0 or
> >> powerful enough to inline some constructs. But this approach had
> >> drawbacks,
> >> especially wrt debug info. These restrictions were gradually lifted and
> >> now
> >> the pragma is entirely piggybacked on the Tree inliner.
> >>
> >> This went mostly OK, except for a few cases where intrisinc operations that
> >> used to be reasonably handled at -O0 now generate awful code, the typical
> >> example being a modulus or division instrinsic by a power-of-2 generating a
> >> fully-fledged modulus or division instruction instead of a simple shift.
> >>
> >> Therefore the attached patch implements anonymous constant propagation in
> >> the
> >> inliner to fix the code quality regression.
> >>
> >> Tested on x86_64-suse-linux, OK for the mainline?
> >>
> >> if (TREE_CODE (*tp) == SSA_NAME)
> >> {
> >> - *tp = remap_ssa_name (*tp, id);
> >> + tree t = remap_ssa_name (*tp, id);
> >> + /* Perform anonymous constant propagation, this makes it possible to
> >> + generate reasonable code even at -O0 for operators implemented as
> >> + inline functions. */
> >> + if (TREE_CODE (t) == SSA_NAME
> >> + && SSA_NAME_DEF_STMT (t)
> >> + && (!SSA_NAME_VAR (t) || DECL_IGNORED_P (SSA_NAME_VAR (t)))
> >> + && gimple_assign_copy_p (SSA_NAME_DEF_STMT (t))
> >> + && is_gimple_min_invariant
> >> + (gimple_assign_rhs1 (SSA_NAME_DEF_STMT (t))))
> >> + *tp = gimple_assign_rhs1 (SSA_NAME_DEF_STMT (t));
> >> + else
> >> + *tp = t;
> >
> > This looks like a good idea to me (though i can't approve it). We may want
> > to
> > lift the (!SSA_NAME_VAR (t) || DECL_IGNORED_P (SSA_NAME_VAR (t))) when
> > optimize
> > is set - the amount of garbage inliner produce is large and killing it
> > early is
> > better than killing it later. This has chance to help early opts where
> > ordering between ccp and einline is quite hard.
>
> Early opts run CCP as pretty much the first pass, so I don't see what
> you are refering to here.
Hmm, you are right. I remember playing with similar patch but that was before
we turned off iteration in early inliner and it was motivated to do more
of indirect call promotion.
Since ipa-prop is no longer complete joke on propagating devirutalization info
perhaps this is no longer too important.
honza
