On Wed, Jul 24, 2019 at 11:29 AM Eric Botcazou <ebotca...@adacore.com> wrote:
>
> Hi,
>
> for EH cleanups, the branch probability heuristics can consider that edges are
> never taken, i.e. their profile count is set to zero. In this case, no matter
> how you tweak the inlining parameters, you cannot get function calls inlined,
> but -Winline nevertheless prints the standard message about unlikely calls as
> in the cases when tweaking the inlining parameters can work.
>
> Therefore the attached patch adds a new CIF code with the warning message:
>
> "call is never executed and code size would grow [-Winline]"
>
> for this case, thus signaling the user that he'd better not waste time trying
> to get the call inlined.
>
> Tested on x86_64-suse-linux, OK for the mainline?
Looks good besides
+ if (e->count.ipa () == profile_count::zero ())
+ e->inline_failed = CIF_NEVER_CALL;
does it actually matter what kind of profile quality e->count.ipa has
compared to profile_count::zero ()? Also
+/* Call is considered never executed. */
+DEFCIFCODE(NEVER_CALL, CIF_FINAL_NORMAL,
+ N_("call is never executed and code size would grow"))
suggests the call is never executed, but we only assume that
(or the profile training run never hit it).
Richard.
>
>
> 2019-07-24 Eric Botcazou <ebotca...@adacore.com>
>
> * cif-code.def (NEVER_CALL): New code.
> * ipa-inline.c (want_inline_small_function_p): Fix formatting issues.
> Set the failure to CIF_NEVER_CALL if the IPA count is zero.
>
> --
> Eric Botcazou
