From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 44765 invoked by alias); 31 Jan 2018 03:46:35 -0000 Mailing-List: contact gdb-patches-help@sourceware.org; run by ezmlm Precedence: bulk List-Id: List-Subscribe: List-Archive: List-Post: List-Help: , Sender: gdb-patches-owner@sourceware.org Received: (qmail 44753 invoked by uid 89); 31 Jan 2018 03:46:34 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=-1.9 required=5.0 tests=AWL,BAYES_00,SPF_HELO_PASS,SPF_PASS,T_RP_MATCHES_RCVD autolearn=ham version=3.3.2 spammy=vision, trading X-HELO: simark.ca Received: from simark.ca (HELO simark.ca) (158.69.221.121) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with ESMTP; Wed, 31 Jan 2018 03:46:33 +0000 Received: by simark.ca (Postfix, from userid 112) id CFA4C1E5B7; Tue, 30 Jan 2018 22:46:31 -0500 (EST) Received: from simark.ca (localhost [127.0.0.1]) by simark.ca (Postfix) with ESMTP id DC4F01E48F; Tue, 30 Jan 2018 22:46:30 -0500 (EST) MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII; format=flowed Content-Transfer-Encoding: 7bit Date: Wed, 31 Jan 2018 03:46:00 -0000 From: Simon Marchi To: Pedro Alves Cc: Yao Qi , gdb-patches@sourceware.org Subject: Re: [PATCH 4/7] Class-fy partial_die_info In-Reply-To: <3655ae02-6fe5-de74-1e4d-f396b200f226@redhat.com> References: <1516873114-7449-1-git-send-email-yao.qi@linaro.org> <1516873114-7449-5-git-send-email-yao.qi@linaro.org> <3655ae02-6fe5-de74-1e4d-f396b200f226@redhat.com> Message-ID: <8f245668d980ad825fd6499c24730f67@simark.ca> X-Sender: simark@simark.ca User-Agent: Roundcube Webmail/1.3.2 X-SW-Source: 2018-01/txt/msg00640.txt.bz2 On 2018-01-30 06:39, Pedro Alves wrote: > On 01/29/2018 01:15 AM, Simon Marchi wrote: > >> From what I understand, the only reason to have that private >> constructor is >> to construct a temporary partial_die_info object used to search in the >> htab, >> is that right? If so, converting that htab_t to an std::unordered_map >> would >> remove the need for all this, since you don't need to construct an >> object >> to search it. See the diff below that applies on top of this patch. >> >> It's not thoroughly tested and I am not sure of the validity of the >> per_cu->cu->partial_dies.empty () call in find_partial_die, but I >> think it >> should work. Plus, it adds some type-safety, which I am a big fan of. >> >> But otherwise, the patch is fine with me. > > Careful here. This could do with some benchmarking. The DWARF reading > code > is performance (both timing and memory) sensitive. This is trading an > open > addressing hash table (htab_t), for a node-based closed addressing hash > table. > The keys/elements in the map are small so I'd expect this to make > a difference. Also, this is trading a in-principle cache-friendly > obstack allocation scheme for the standard new allocator. Ah, indeed. I thought that unordered_map would be implemented the same way as htab_t, but I see it's not the case. Doing some quick tests on a big binary, it increases the time reading the symbols from an average of 37 seconds to an average of 42 seconds. I understand the different hash table implementation having an impact, but I don't really understand how the allocation scheme can have a meaningful impact. The partial_die_info objects are still allocated on the obstack, aren't they? So it's just the space for the table itself that isn't on the objstack, but I don't see why that would make a difference. If we want to have a data structure with the same kind of performance as htab_t but with type-safety in the future, is your vision that we'll have to implement it ourselves? Should we make a wrapper around htab_t? Simon