From: Al Viro Subject: Re: sparse -Wptr-subtraction-blows: still needed? Date: Wed, 2 May 2007 01:35:57 +0100 Message-ID: <20070502003557.GK4095@ftp.linux.org.uk> References: <4637AC3D.6000008@freedesktop.org> <4637D47D.7040203@freedesktop.org> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from zeniv.linux.org.uk ([195.92.253.2]:44158 "EHLO ZenIV.linux.org.uk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S2992436AbXEBAgB (ORCPT ); Tue, 1 May 2007 20:36:01 -0400 Content-Disposition: inline In-Reply-To: Sender: linux-sparse-owner@vger.kernel.org List-Id: linux-sparse@vger.kernel.org To: Linus Torvalds Cc: Josh Triplett , linux-sparse@vger.kernel.org, linux-kernel@vger.kernel.org On Tue, May 01, 2007 at 05:24:54PM -0700, Linus Torvalds wrote: > The only thing that was gcc-specific about it is that gcc goes to some > extreme lengths to turn the constant-sized division into a sequence of > shifts/multiples/subtracts, and can often turn a division into something > like ten cheaper operations instead. > > But that optimization was also what made gcc take such a long time if the > constant division is very common (ie a header file with an inline > function, whether that function is actually _used_ or not apparently > didn't matter to gcc) There used to be a Dumb Algorithm(tm) in there - basically, gcc went through all decompositions of constant it was multiplying at and it did it in the dumbest way possible (== without keeping track of the things like "oh, we'd already looked at the best way to multiply by 69, no need to redo that again and again)"). _That_ is what blew the build time to hell for a while. And yes, that particular idiocy got fixed (they added a moderately-sized LRU of triplets , which killed recalculation from scratch for each pointer subtraction and got complexity of dealing with individual subtraction into tolerable range). That's separate from runtime issues, of course.