From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-wm1-x343.google.com (mail-wm1-x343.google.com [IPv6:2a00:1450:4864:20::343]) by sourceware.org (Postfix) with ESMTPS id 63450393C877 for ; Wed, 30 Sep 2020 10:18:35 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org 63450393C877 Authentication-Results: sourceware.org; dmarc=none (p=none dis=none) header.from=mittosystems.com Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=jozef.l@mittosystems.com Received: by mail-wm1-x343.google.com with SMTP id e2so1123544wme.1 for ; Wed, 30 Sep 2020 03:18:35 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=mittosystems.com; s=google; h=date:from:to:cc:subject:message-id:mail-followup-to:references :mime-version:content-disposition:in-reply-to; bh=rGv3QZZFm0eL1zwJv6gHVWsOcdaKLgliEXZJKE/MND0=; b=e2WdAZTp4jbUrkwxZdP8Kp8OUhbFRgngEwQmqPUE5Nuu2kHjH5EpU7hW7MlKQjCVnN SxJKAhkbSfCo3o/TlgPyA0Z5CLVPL7lJPVhBrJpesE3pEbdM6zFH4qqlZtum1Y753f3M V9EqmDiFzftHUbPJLrEq0qi8jZtFuZUSHZ5eJMJKrVr53FQp6EU0zFKw5HsgO/+3P9Vb FwBImXOSi+OG9ySRT2wWcL03YoNuhp8dVMY4MNrOcFENafR0GOq0SkfHB60RzlaMOBq0 EIyQm2XjF/38msLuB1fbdS7nA4foLgetkgvJtCIngwX7yF9/Na2rTCsfvcLIbC40w1qL NxRA== X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:date:from:to:cc:subject:message-id :mail-followup-to:references:mime-version:content-disposition :in-reply-to; bh=rGv3QZZFm0eL1zwJv6gHVWsOcdaKLgliEXZJKE/MND0=; b=B59FLgjgD/PA8CSy+UjJK0ynKxOCsktvQ9re37khXdS/Atw7KmC/h3pHrALVCOdnKl +9NU62OAMecP1jUjQmpp9bf7ZaV+FT7sfkUiuvJkAVeh9UExDsCA2Ne0qRLKRIBioEPq 7A/NXwgUi4rFyQloVjvBiWV18c/zB5RzLLu1Fd8WqaFuByyA7noxNbpBTQTbCzSJoqZj DISuBTz/Wc+MK6ACsnEUSV/ynr5kq4kwVu5/dC43oXO8cRnFCbUgMywEj6gsMqS/u4Ha JNsSSw7keV/9gxwo7Xu3arX0cxoovNXFmRSQ+wMxQBA2yeENtu6zCa5dCKpYPhW4+akC I2vg== X-Gm-Message-State: AOAM533MSC6y4VD48Uzup2zSJ/OPWcUcD8clGxPc5gDY47G1Z+0cqkYc ay8SQPhif2B+ZkrTQdslidLjEg== X-Google-Smtp-Source: ABdhPJxhnANcJEwN2qgMnwXt2RGPMJDfeaZFfx8TXQ3G8wKq+wqaG0dbxNEBIoK3J5tLz1Q2aMq2HQ== X-Received: by 2002:a05:600c:216:: with SMTP id 22mr2080822wmi.149.1601461113860; Wed, 30 Sep 2020 03:18:33 -0700 (PDT) Received: from jozef-acer-manjaro ([2a01:4b00:87fd:900:5e1d:5c99:56da:76e8]) by smtp.gmail.com with ESMTPSA id q4sm2065101wru.65.2020.09.30.03.18.32 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Wed, 30 Sep 2020 03:18:33 -0700 (PDT) Date: Wed, 30 Sep 2020 11:18:31 +0100 From: Jozef Lawrynowicz To: Roland McGrath Cc: Fangrui Song , Binutils Development , gnu-gabi@sourceware.org Subject: Re: [PATCH v2] Support for SHF_GNU_RETAIN ELF Section Flag Message-ID: <20200930101831.vi7pxhstftemanfc@jozef-acer-manjaro> Mail-Followup-To: Roland McGrath , Fangrui Song , Binutils Development , gnu-gabi@sourceware.org References: <20200928132613.btkqaoomv4fdnupn@jozef-acer-manjaro> <20200929044353.vx26ypc2oioqbmfb@gmail.com> <20200929100442.o3e66jbmqpv5cajh@jozef-acer-manjaro> <20200929193806.m6u6ra6oijqkstfo@gmail.com> <20200929213741.jqegh62d7jne5uyo@jozef-acer-manjaro> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: X-Spam-Status: No, score=-3.0 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP autolearn=ham autolearn_force=no version=3.4.2 X-Spam-Checker-Version: SpamAssassin 3.4.2 (2018-09-13) on server2.sourceware.org X-BeenThere: gnu-gabi@sourceware.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Gnu-gabi mailing list List-Unsubscribe: , List-Archive: List-Help: List-Subscribe: , X-List-Received-Date: Wed, 30 Sep 2020 10:18:38 -0000 On Tue, Sep 29, 2020 at 05:10:09PM -0700, Roland McGrath wrote: > Since group semantics mean that if any section in a group is retained the > whole group must be retained, it seems to me that making the feature a > group flag rather than a section flag makes sense. I don't think it really > has any downside. It's a little bit more structure to generate a group but > everything already knows how to do it so that should be no problem. While > it is true that multiple sections with the same name and different > attributes should already be fine, the place where multiple different > sections of the same name are usually seen today is when each different > section reusing a name is in a different group from the other sections with > the same name. > Let's assume: GRP_RETAIN Sections in a group with this flag set should always be included in the linked object, even if they appear unused. One issue I have with the group method is conceptual; groups are supposed to be used to group related sections. They are related because of their meaning from the application perspective, not because of their metadata (i.e. sections to be retained are related by their metadata because they all should be retained). Just because groups have a side effect that can be leveraged to get a section to be retained in the final link, why does that mean we need to leverage that property when it it's not actually the simplest or most obvious way to implement the new behavior? SHF_GNU_RETAIN describes in the simplest possible way that the section should be retained in the final link, without relying on other constructs. Just because groups are discarded or retained as a group, doesn't mean we need to leverage groups to try and implement retention or discarding functionality. Why wasn't SHF_EXCLUDE implemented as a group flag? After all, groups are included or excluded from the link together. GRP_EXCLUDE Sections in a group with GRP_EXCLUDE set should be discarded from the link. I mean, you could kind of use groups for anything when you decide grouping sections by metadata is OK. Why define SHT_NOBITS when you can create: GRP_NOBITS Sections in this group occupy no space in the file. They must have type SHT_PROGBITS. Retention of a section is a property of the section. We are misusing ELF constructs by using groups to indicate an arbitrary section needs to be retained. Another issue (if more is needed) is about how to name the groups. * If it is mandated that GRP_RETAIN groups have the same name e.g. "grp_retain", that means you can't put a section you want to retain in a different logical group that makes sense from the application perspective. So the other sections you would want to put in a group with the retained section need to all be put in a bundle with all the other GRP_RETAIN sections. * If GRP_RETAIN groups can have any name, so you can have multiple GRP_RETAIN groups, how does the compiler decide how to name the groups? It seems like it would be a mess. "grp_retain0", "grp_retain2" ... "grp_retain10" from one object file, "grp_retain0"... "grp_retain5" from another. Extra processing in the linker to clean up these group names and keep them unique would be required when performing a relocatable link. As a general point, what if I decide that there's enough pressure from the anti-SHF_GNU_RETAIN side that I change the implementation to use groups. But then those who already backed the flag prefer that method over using groups think the implementation should not have been changed. I feel like we already got enough backing for SHF_GNU_RETAIN between the GNU gABI and Binutils discussions. I understand, and welcome, more feedback, but I haven't been convinced any other method is obviously better than SHF_GNU_RETAIN, so why change it? I'm not saying it's possible to quantify which mechanism for "retain" is best, but SHF_GNU_RETAIN is the simplest, most obvious, and easiest to understand. Surely that gives it top marks? I'm also yet to hear one convincing reason why SHF_GNU_RETAIN is bad. As far as I can tell, the main argument against it stems from the fact that you can have two input sections with the same name but different SHF_GNU_RETAIN flag states. Not only is this a non-issue, groups have exactly the same problem! It would be great if a global maintainer to chime in on whether the attached patch is acceptable, because otherwise we are going to go back and forth forever. Thanks, Jozef > On Tue, Sep 29, 2020 at 2:37 PM Jozef Lawrynowicz > wrote: > > > On Tue, Sep 29, 2020 at 12:38:06PM -0700, Fangrui Song wrote: > > > On 2020-09-29, Jozef Lawrynowicz wrote: > > > > There are more logic holes in your other email that I didn't > > > > respond to, but I've already spent a tonne of time trying to address > > > > your points (most of which you ignored, so why did I bother?), so I > > > > think I'm done debating this. > > > > If other people back up your suggestion, and address my concerns, then > > > > we can re-open this. > > > > > > Hi Jozef, apologies if I did not respond to your previous replies in > > > time or if I acted as if I "blocked" your change. I maintain LLVM's > > > linker LLD and am also a contributor to its binary utilities > > > (llvm-readelf, llvm-objdump, ...). If GCC makes use of this flag for > > __attribute__((used)), > > > LLD and LLVM's binary utilities do need to support it. > > > So I want to participate in the design before we regret :) > > > > > > I have studied linker GC quite a lot (contributed patches in LLD > > > and coordinated on the feature parity among the 3 linkers > > > (e.g. https://sourceware.org/pipermail/binutils/2020-August/112732.html) > > > > > > LLD and GNU ld, LLVM's integrated assembler and GNU as have achieved > > consensus on many aspects > > > on SHF_LINK_ORDER and 'unique' linkage. Credits to H.J. for all the > > binutils side work. > > > ) > > > > > > I have asked a couple of toolchain folks. Many do share the same feeling > > > with me: an additional GC feature is useful but whether it should be > > > this exact form of a new flag we don't know. > > > > Hi Fangrui, > > > > Thank you for providing some more information to backup your points, I > > will address those below. > > > > > > > > Your patch also deprives an OS-specific flag from ELFOSABI_FREEBSD:) > > > I am not a committer of FreeBSD and I am surely not authoritative on > > > whether this is acceptable. But, FreeBSD does use the aforementioned > > > LLVM tools and I don't want to dirty the code base by adding > > > if (osabi == ELFOSABI_FREEBSD) do something. > > > > I don't have any insight into the ELFOSABI_FREEBSD change, H.J. added that > > part, I assume it is very closely aligned with the "pure" GNU OSABI, so > > any changes to one should be kept in parity with the other. However, > > that is only my assumption. > > > > > > > > We have 8 bits in SHF_MASKOS: not too few, but also not too many. We > > > should be prudent with new flags, especially if we already have a > > > facility for section retaining. There is a difference ".reloc requires a > > > known live section" which might cause friction but it has the advantage > > that it is > > > time-tested and has good toolchain support. > > > > Yes, this is something I considered, but there are 6 bits remaining to > > be used, with SHF_GNU_RETAIN there will be 5 bits remaining. > > Given how little movement there has been within the SHF_MAKOS bits for > > the past 20 years, I don't think we need to be too protective of the > > bits. They are there to be used after all, and IMO SHF_GNU_RETAIN is a > > very well-defined use case to use one of those bits. > > > > For the .reloc mechanism, in every object file with a retained section > > you will need to have an empty input section which is guaranteed to be > > kept by the linker. How is the compiler to know what that section is? > > > > > > > > > > > Opinions from others: > > > > > > * Peter Smith mentioned that Arm's proprietary toolchain does something > > > similar to SHF_GNU_RETAIN > > > ( > > https://www.keil.com/support/man/docs/armcc/armcc_chr1359124983230.htm) > > > "I put __attribute__((used)) on my function and yet the tools removed > > > it anyway." > > > > > > * James Henderson mentioned whether we can > > > "create a new section type listing section indexes that should be > > > retained" > > > As a variant, we could use a SHT_GROUP with a new GRP_* bit. > > > Anchor: [group] (see below) > > > > Using groups was discussed on the GNU gABI thread, but I don't really > > see the benefit of the added complication. Doesn't handling new section > > group types still require significant coordination between all the > > different compilers, assemblers and linkers? > > > > Any attempt to reference a section which does not exist in the object > > file, and is not guaranteed to be present in the linked output file, is > > a non-starter in my opinion. We give users the freedom to do what they > > want with linker scripts as ELF doesn't mandate any section must be > > linked into the final executable. I think it's better to consolidate > > the requirement to retain within the section that requires retaining, > > instead of spreading the requirement to retain between ELF constructs > > which have unrelated purposes, and need to be aligned between different > > stages of the build process. > > > > > > > > > Do we want to make life easier for ourselves, or easier for our users? > > > > > > > > I get that ABI changes can be a bit disruptive, but this new flag in > > particular really isn't complicated anyway. > > > > > > On 2020-09-29, Jozef Lawrynowicz wrote: > > > > On Mon, Sep 28, 2020 at 09:43:53PM -0700, Fangrui Song wrote: > > > > > On 2020-09-28, Jozef Lawrynowicz wrote: > > > > > > The attached patch is version 2 of the SHF_GNU_RETAIN patch that > > was > > > > > > previously discussed here: > > > > > > > > https://sourceware.org/pipermail/binutils/2020-September/113406.html > > > > > > > > > > > > The following changes have been made: > > > > > > - Removed the .retain directive > > > > > > - The assembler will create different input sections for sections > > with > > > > > > the same name but SHF_GNU_RETAIN set/unset (thanks to H.J. for > > that). > > > > > > This means the linker will be able to do a better job garbage > > > > > > collecting input sections, as the "retain" attribute applied to a > > > > > > symbol declaration in the source code will not cause other parts > > of > > > > > > the program that are not required, but are in the same section, > > to be > > > > > > unnecessarily retained. > > > > > > - Added GNU OSABI handling (also thanks to H.J.). > > > > > > > > > > My point from > > https://sourceware.org/pipermail/binutils/2020-September/113466.html > > stands. > > > > > Section flags are a bit cumbersome. If the following > > > > > > > > > > // a.h > > > > > __attribute__((section("sec"))) > > > > > inline void bar() { ... } > > > > > // a.c > > > > > #include "a.h" > > > > > __attribute__((section("sec"), retain)) > > > > > void foo() { > > > > > } > > > > > > > > > > compiles to > > > > > > > > > > .section sec,"a",@progbits > > > > > ... > > > > > .section sec,"aR",@progbits > > > > > ... > > > > > > > > > > You will get a gas error for changing section flags > > > > > (https://sourceware.org/pipermail/binutils/2020-February/109945.html > > ) > > > > > > > > > > > > > There is no error in this case. > > > > The original patch handled it, and it has been further improved in v2. > > > > There's even a testcase testing exactly the functionality you say is > > > > broken. > > > > > > > > gas/testsuite/gas/elf/section22.s: > > > > .section .bss,"aw" > > > > ... > > > > .section .bss,"awR",%nobits > > > > > > > > I already gave a detailed response to your reloc proposal: > > > > https://sourceware.org/pipermail/binutils/2020-September/113450.html > > > > > > Letting the following assembly creates two .bss sections adds an > > exception to > > > "gas: error for section type, attr, or entsize change" > > > (https://sourceware.org/pipermail/binutils/2020-February/109945.html) > > > for which I have achieved consensus with Alan Modra and some LLVM > > > toolchain folks. That is why I hope we could re-consider this design > > > choice. > > > > > > .section .bss,"aw" > > > .section .bss,"awR",%nobits > > > > I agree that this could be considered slightly inelegant, but much less > > so than having .reloc directives in arbitrary empty sections. > > Once someone understands what the "R" flag does, it should be clear why > > we want to create a separate input section for it. > > > > If we really want, we can force the compiler to create unique section > > names for declarations which have the "retain" attribute applied. I > > assume most of the time, users will want this fine section > > granularity. I would prefer not to go that route and leave it to the > > users to specify unique sections if they want. > > > > > > > > See above for [group]: it is actually nice to use a group here, because > > > the quadruple > > > (section name, group name if "G", linked-to section if "o", unique id of > > > "unique") is the section key. The following creates a section different > > > from the default .bss: > > > > > > .section .bss,"awG",%nobits,gc_root > > > > I don't think it is really very important whether input sections with > > the same name are unique or not within an object file. ELF doesn't care > > about this. You can have sections with the same name and same flags in > > different object files and the linker can just deal with it. It's the > > same situation if you have two sections from the same object file which > > have the same name. > > > > Within an object file, you'll end up with all the non-SHF_GNU_RETAIN > > sections with the same name being merged together by the assembler, and > > all those with SHF_GNU_RETAIN being merged. I don't think there is any > > benefit to keep SHF_GNU_RETAIN sections with the same name as separate > > input sections in the object file. > > > > Thanks, > > Jozef > > > > > > > > > Why don't you first answer my questions about why using relocs is a > > > > really bad idea, is a hack, doesn't conceptually make any sense, and > > > > will confuse users? > > > > > > > > You never actually gave a reason that using relocs is better than > > > > SHF_GNU_RETAIN, you seem to be pushing this alternative implementation > > just > > > > because it is an alternative implementation. I don't count your > > > > description of section flags as "cumbersome" as a reason we should > > > > shelve this proposal. > > > > > > I hope my previous paragraphs have answered this point. > > > > > > > > .reloc is really convenience in this case. You can add as many .reloc > > > > > directives as you like, each contributing one R_*_NONE to the object > > > > > file but zero cost to the linked image. > > > > > > > > Are we really counting a section flag as adding "cost" to a linked > > > > image? Your method requires a "benign zero-sized section" (suggested > > in > > > > your previous email), doesn't this also add cost? What about the cost > > of > > > > having nonsensical relocs in object files? > > > > > > > > If we are using the term "cost" loosely, and to further our own point, > > > > then SHF_GNU_RETAIN actually has negative cost to the linked image. You > > > > can have different input sections with the same name, but with > > different > > > > states for SHF_GNU_RETAIN (set and unset), and so garbage collection > > > > will operate with finer granularity, removing more unneeded parts of > > the > > > > program compared to your reloc mechanism, which would keep the entire > > > > section containing the symbol referenced by the reloc. > > > > > > > > Thanks, > > > > Jozef > > > >