Re: [PATCH] Add XINSN macro and use it within NEXT_INSN/PREV_INSN (was Re: [PATCH] Force rtl templates to be inlined)

[David Malcolm <dmalcolm@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2235 bytes --]

On Fri, 2014-09-05 at 12:45 -0600, Jeff Law wrote:
> On 09/04/14 14:04, David Malcolm wrote:
> > On Tue, 2014-09-02 at 19:50 +0200, Andi Kleen wrote:
> >>> I suspect the bulk of them currently are coming from the safe_as_a
> >>> <rtx_insn *> calls within NEXT_INSN and PREV_INSN; do you happen to have
> >>> information handy on that?
> >>
> >> Yes that's right:
> >>
> >> -   1.03%  lto1                    [.] bool is_a_helper<rtx_insn*>::test<rtx_def>(rtx_def*)                                                     â–’
> >>     - bool is_a_helper<rtx_insn*>::test<rtx_def>(rtx_def*)                                                                                       â–’
> >>        - 92.20% bool is_a<rtx_insn*, rtx_def>(rtx_def*)                                                                                          â–’
> >>           - 98.53% rtx_insn* safe_as_a<rtx_insn*, rtx_def>(rtx_def*)                                                                             â–’
> >>              - 73.28% NEXT_INSN(rtx_insn const*)                                                                                                 â–’
> >
> > The is_a_helper for rtx_insn * is non-trivial, so it may be worth
> > avoiding it, even when inlined.
> >
> > The attached patch rewrites the inline NEXT_INSN/PREV_INSN to avoid
> > doing the safe_as_a, instead tightening up the interface so that one can
> > only set them to an insn, and introducing a new XINSN access macro and
> > corresponding rt_insn member of the union.
> >
> > Bootstrapped on x86_64-unknown-linux-gnu (Fedora 20), and has been
> > rebuilt as part of a config-list.mk build for all working configurations
> > (albeit with other patches for the latter case).
> >
> > OK for trunk?
> So is this just to deal with the overhead in the safe_as_a helper until 
> we can strengthen more code?  And is that overhead significant in an 
> optimized build?
> 
> Would an alternate approach be to make the checking in safe_as_a 
> conditionalized on ENABLE_CHECKING?

I wrote the attached patch to do that, but I don't yet have numbers to
back it up.

Bootstrapped with current settings, and smoketested with
--enable-checking=release (both on Fedora 20 x86_64).

[-- Attachment #2: 0001-Ensure-that-safe_as_a-can-have-no-performance-overhe.patch --]
[-- Type: text/x-patch, Size: 960 bytes --]

From 24b98616c4425a8ef380d2d5fef00f82af9df985 Mon Sep 17 00:00:00 2001
From: David Malcolm <dmalcolm@redhat.com>
Date: Fri, 29 Aug 2014 17:04:49 -0400
Subject: [PATCH] Ensure that safe_as_a can have no performance overhead in a
 release build

gcc/
	* is-a.h (safe_as_a): Eliminate the conditional from the
	!defined(ENABLE_CHECKING) case, to ensure there can be no
	performance degradation of the release build.
---
 gcc/is-a.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/gcc/is-a.h b/gcc/is-a.h
index 176066b..f0c05b0 100644
--- a/gcc/is-a.h
+++ b/gcc/is-a.h
@@ -200,6 +200,7 @@ template <typename T, typename U>
 inline T
 safe_as_a (U *p)
 {
+#ifdef ENABLE_CHECKING
   if (p)
     {
       gcc_checking_assert (is_a <T> (p));
@@ -207,6 +208,9 @@ safe_as_a (U *p)
     }
   else
     return NULL;
+#else
+  return is_a_helper <T>::cast (p);
+#endif
 }
 
 /* A generic checked conversion from a base type U to a derived type T.  See
-- 
1.8.5.3