Re: c++/1687: Extreme compile time regression from 2.95.2

["Zack Weinberg" <zackw@Stanford.EDU>]

The following reply was made to PR c++/1687; it has been noted by GNATS.

From: "Zack Weinberg" <zackw@Stanford.EDU>
To: Scott A Crosby <crosby@qwes.math.cmu.edu>
Cc: Kelley Cook <kelley.cook@home.com>, gcc-gnats@gcc.gnu.org,
   gcc-bugs@gcc.gnu.org
Subject: Re: c++/1687: Extreme compile time regression from 2.95.2
Date: Wed, 14 Feb 2001 13:00:49 -0800

 On Wed, Feb 14, 2001 at 06:07:08AM -0500, Scott A Crosby wrote:
 > On Tue, 6 Feb 2001, Zack Weinberg wrote:
 > 
 > > Using the 'visit each node only once' mechanism of walk_tree
 > > thoroughly squelches the performance problem.  (One can't use
 > > walk_tree_without_duplicates blindly - slightly more cleverness is
 > > required.  It's still simple.)  We get sub-second compile time all the
 > > way up to -O2 and 32 input mux().
 > 
 > > HOWEVER: I am not certain that the change is correct.  Suppose that a
 > > function A is a candidate for inlining, and it's called twice from the
 > > same function B.  If the two call_expr nodes are shared, we won't
 > > inline both calls.  There may be other problems as well.
 > > 
 > > Patch follows.  Commentary from C++ team appreciated.  Will bootstrap
 > > and report.
 > 
 > 
 > Could you workaround this by walking the tree normally, and try to inline
 > at all of the nodes, but if you find out that you are scanning too many
 > nodes, you abort and use a workaround strategy, say, walk-each-node once?
 > 
 > Catch the pathalogical cases and shunt them elsewhere, and behave normally
 > otherwise?
 
 Let's find out if always walking each node once is safe, first.  It
 doesn't cause any C++ regressions, but I still don't know if it
 inhibits optimizations.  Paging C++ team...
 
 zw