From mboxrd@z Thu Jan 1 00:00:00 1970 From: Luc Van Oostenryck Subject: Re: ptrlist-iterator performance on one wine source file Date: Sun, 30 Jul 2017 17:12:55 +0200 Message-ID: <20170730151254.xlnz7c4zphhnhump@ltop.local> References: <20170729130116.b5jd4rvkiwzgsfwt@ltop.local> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from mail-wm0-f41.google.com ([74.125.82.41]:35284 "EHLO mail-wm0-f41.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750817AbdG3PNA (ORCPT ); Sun, 30 Jul 2017 11:13:00 -0400 Received: by mail-wm0-f41.google.com with SMTP id m85so56541067wma.0 for ; Sun, 30 Jul 2017 08:12:59 -0700 (PDT) Content-Disposition: inline In-Reply-To: Sender: linux-sparse-owner@vger.kernel.org List-Id: linux-sparse@vger.kernel.org To: Christopher Li Cc: Linux-Sparse , Dibyendu Majumdar On Sun, Jul 30, 2017 at 12:15:40AM -0400, Christopher Li wrote: > On Sat, Jul 29, 2017 at 5:47 PM, Luc Van Oostenryck > wrote: > > > > We do, more or less. > > Once the code is linearized and inlining is done we: > > - never add new BBs > > - remove some BBs (some removing edges from the CFG) > > - do several kinds of branch simplification (that's moving > > edge, so technically it's adding edge to the CFG, not sure it > > change the dom tree though). > > > That is merging nodes right? Two nodes combine as one. I was more thinking to things like done in try_to_simplify_bb(). > Moving block to another place is another story. > > > > Yes, but this calculation is not correct at all. > > - each time a node is removed, the total number of nodes is smaller > > and so the next time is a bit faster (this would correspond to a factor > > of 2) > > if N >> M then it does not matter much. Indeed. > > - much more importantly, each time kill_unreachable_bbs() is called > > *all* the currently dead BBs are removed at once. So single call > > can kill several BBs. Of course, it will be different for each CFG/input > > files. > > Yes, that would be linear to the number of blocks removed. It still > need to go through the blocks to clean up the instructions usage etc. Yes, but that's totally independent of how and how often the detection of dead code is done. At the end, every dead instructions will need to be cleaned up (the minimum is removing pseudo usage). > >> In the memops finding dominating store is doing a lot worse. That is > >> why gcc complete that file almost instantly. Sparse takes 30 seconds > >> on my machine. One big problem is it did not cache the dominating > >> result. It is redoing the finding again and again. > > > Uh? > > Which input file your talking about? > > This ptrlist testing wine source file that takes 23 second for sparse to run. > I take a brief look at it, it is doing a lot of dominating search. Is it possible to have a pathname or a link? > > *smile* Feels like linear? > > Did you try with several input files, some with big functions? > I just try some sparse source file. The largest one is in parse.c. It's not the size of the file that matter here, it's the size (and complexity) of the function(s). > The paper has more detail report on using huge number of nodes. > Tested on real code and random generated CFG for really large > number of nodes. I am not going repeat it here. No, I know about this paper. > BTW, I just find out LLVM was using the exact same algorithm at > some point. Not sue what they are using now. They might not build > the whole tree any more. It's possible, indeed. -- Luc