<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 13, 2016 at 9:15 PM, Hal Finkel <span dir="ltr"><<a href="mailto:hfinkel@anl.gov" target="_blank">hfinkel@anl.gov</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div style="font-family:arial,helvetica,sans-serif;font-size:10pt;color:#000000"><br><br><hr id="m_-1640572931879209436m_2604153376224275460zwchr"><blockquote style="border-left:2px solid rgb(16,16,255);margin-left:5px;padding-left:5px;color:rgb(0,0,0);font-weight:normal;font-style:normal;text-decoration:none;font-family:Helvetica,Arial,sans-serif;font-size:12pt"><b>From: </b>"Sean Silva via llvm-dev" <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>><br><b>To: </b>"Rui Ueyama" <<a href="mailto:ruiu@google.com" target="_blank">ruiu@google.com</a>><br><b>Cc: </b>"llvm-dev" <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>><br><b>Sent: </b>Tuesday, December 13, 2016 9:59:37 PM<span><br><b>Subject: </b>Re: [llvm-dev] LLD status update and performance chart<br><br></span><div dir="ltr"><br><div class="gmail_extra"><br><div><div class="m_-1640572931879209436h5"><div class="gmail_quote">On Tue, Dec 13, 2016 at 12:08 PM, Rui Ueyama <span dir="ltr"><<a href="mailto:ruiu@google.com" target="_blank">ruiu@google.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><div class="m_-1640572931879209436m_2604153376224275460gmail-h5">On Tue, Dec 13, 2016 at 12:01 PM, Mehdi Amini <span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div style="word-wrap:break-word"><br><div><div><div class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107h5"><blockquote><div>On Dec 13, 2016, at 11:51 AM, Rui Ueyama <<a href="mailto:ruiu@google.com" target="_blank">ruiu@google.com</a>> wrote:</div><br class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726Apple-interchange-newline"><div><div dir="ltr" style="font-family:helvetica;font-size:12px;font-style:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 13, 2016 at 11:37 AM, Mehdi Amini<span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726Apple-converted-space"> </span><span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span><span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726Apple-converted-space"> </span><wbr>wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div style="word-wrap:break-word"><br><div><div><div class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726h5"><blockquote><div>On Dec 13, 2016, at 11:30 AM, Rui Ueyama <<a href="mailto:ruiu@google.com" target="_blank">ruiu@google.com</a>> wrote:</div><br class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866Apple-interchange-newline"><div><div dir="ltr" style="font-family:helvetica;font-size:12px;font-style:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 13, 2016 at 11:23 AM, Mehdi Amini<span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866Apple-converted-space"> </span><span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span><span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866Apple-converted-space"> </span><wbr>wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div style="word-wrap:break-word"><br><div><span><blockquote><div>On Dec 13, 2016, at 11:08 AM, Rui Ueyama <<a href="mailto:ruiu@google.com" target="_blank">ruiu@google.com</a>> wrote:</div><br class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866m_1491349387806721298Apple-interchange-newline"><div><div dir="ltr" style="font-family:helvetica;font-size:12px;font-style:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 13, 2016 at 11:02 AM, Mehdi Amini<span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866m_1491349387806721298Apple-converted-space"> </span><span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span><span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866m_1491349387806721298Apple-converted-space"> </span><wbr>wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div style="word-wrap:break-word"><div><div class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866m_1491349387806721298gmail-h5"><br><div><blockquote><div>On Dec 13, 2016, at 10:06 AM, Rui Ueyama <<a href="mailto:ruiu@google.com" target="_blank">ruiu@google.com</a>> wrote:</div><br class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866m_1491349387806721298gmail-m_-656514987314485984Apple-interchange-newline"><div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 13, 2016 at 9:28 AM, Mehdi Amini<span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866m_1491349387806721298Apple-converted-space"> </span><span dir="ltr"><<a href="mailto:mehdi.amini@apple.com" target="_blank">mehdi.amini@apple.com</a>></span><span class="m_-1640572931879209436m_2604153376224275460gmail-m_2522788196438720107m_2989931454347399726m_5610684116206409866m_1491349387806721298Apple-converted-space"> </span><wbr>wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><span><br>> On Dec 13, 2016, at 5:55 AM, Rafael Avila de Espindola via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:<br>><br>> Sean Silva via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> writes:<br>>> This will also greatly facilitate certain measurements I'd like to do<br>>> w.r.t. different strategies for avoiding memory costs for input files (esp.<br>>> minor faults and dTLB costs). I've almost gotten to the point of<br>>> implementing this just to do those measurements.<br>><br>> If you do please keep it local. The bare minimum we have of library<br>> support is already disproportionately painful and prevents easier sharing<br>> with COFF. We should really not add more until the linker is done.<br><br></span>This is so much in contrast with the LLVM development, I find it quite hard to see this as an acceptable position on llvm-dev.<br></blockquote><div><br></div><div>LLD is a subproject of the LLVM project, but as a product, LLD itself is not LLVM nor Clang, so some technical decisions that make sense to them are not directly be applicable or even inappropriate. As a person who spent almost two years on the old LLD and 1.5 years on the new LLD, I can say that Rafael's stance on focusing on making a good linker first really makes sense. I can easily imagine that if we didn't focus on that, we couldn't make this much progress over the past 1.5 year and would be stagnated at a very basic level. Do you know if I'm a person who worked really hard on the old (and probably "modular" whatever it means) linker so hard? I'm speaking based on the experience. If you have an concrete idea how to construct a linker from smaller modules, please tell me. I still don't get what you want. We can discuss concrete proposals, but "making it (more) modular" is too vague and not really a proposal, so it cannot be a productive discussion.</div><div><br></div><div>That said, I think the current our "API" to allow users call our linker's main function hit the sweet spot. I know at least a few LLVM-based language developers who want to eliminate external dependencies and embed a linker to their compilers. That's a reasonable usage, and I think allowing them to pass a map from filename to MemoryBuffer objects makes sense, too. That would be done without affecting the overall linker architecture. I don't oppose to that idea, and if someone wrote a patch, I'm fine with that.</div></div></div></div></div></blockquote></div><div><br></div></div></div><div>I’m totally willing to believe you that it is not possible to write the fastest ELF linker on earth (or in the universe) with a library based and reusable components approach. But clang is not the fastest C/C++ compiler available, and LLVM is not the fastest compiler framework either!</div><div><br></div><div>So as a project, it seems to me that LLVM has not put the tradeoff on the speed/efficiency historically when it was to the detriment of layering/component/modularity/<wbr>reusability/…</div><div><br></div><div>Writing the fastest linker possible is nice goal, I regret that a LLVM subproject is putting this goal above layering/component/modularity/<wbr>reusability/… though.</div></div></blockquote><div><br></div><div>I've never mentioned that creating the fastest linker is the only goal.</div></div></div></div></div></blockquote><div><br></div></span><div>I believe this has clearly been put *ahead* the other design aspects I mentioned, isn’t it?</div><span><div><br></div><blockquote><div><div dir="ltr" style="font-family:helvetica;font-size:12px;font-style:normal;font-weight:normal;letter-spacing:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><div class="gmail_extra"><div class="gmail_quote"><div>Medhi, please tell how you would *actually* layer linkers with fine-grained components.</div></div></div></div></div></blockquote><div><br></div></span><div>That’s not a bait I’m gonna bite.</div></div></div></blockquote><div><br></div><div>That's not a bait... I guess you are proposing a different architecture, so you need to explain it.</div></div></div></div></div></blockquote><div><br></div></div></div><div>That’s a bait in the sense that I’m not having two months to dig into the Elf lld and write a design document/proposal for the sole purpose of making a point. And me not doing this does not impact and is not relevant to the discussion at stance..</div></div></div></blockquote><div><br></div><div>If it takes two months for you to investigate and make a proposal, why are you so confident about the conclusion of the investigation that what you think (I still don't get what it is) is doable? I actually dug into the old LLD for two years (not months) with the hope that there's a good way of make it work but failed.</div></div></div></div></div></blockquote><div><br></div></div></div><div>You can’t be serious. You’re bait is asking for me proposing a different architecture. I’m not biting.</div><div>This is *not* the baseline. For instance, you said from the start that you won’t return an Error from APIs and instead call exit(). I don’t need to propose a “linker architecture” for that.</div><div>The main contention point is how any library design guidelines is rejected from the start on the principle that it’ll slow down the linker.</div></div></div></blockquote><div><br></div></div></div><div>That's simply not true as you know. I listened to you, and it now returns. And even in this thread, I mentioned that embedding a linker is a reasonable usage.</div></div></div></div>
</blockquote></div><br></div></div></div><div><div class="m_-1640572931879209436h5"><div class="gmail_extra">Yes. Rui has bent over backwards every time a real user has come to us and said "we need X". The historical precedent here is that LLD is open to many kinds of changes, but not on theoretical grounds.</div><div class="gmail_extra"><br></div><div class="gmail_extra">Admittedly this leads to a somewhat conservative design for the linker w.r.t. enabling new use cases. However, having been involved directly or indirectly with LLD development over the last 4 or 5 years, the number of new use cases that have been proposed fall into a very small number of categories:</div><div class="gmail_extra"><br></div><div class="gmail_extra">- I want to have "main() in a library" for a static linker. LLD/ELF currently provides this functionality (with some unfortunate caveats w.r.t. re-entrancy and other stuff, but the core functionality is there, and LLD has catered to reasonable requests to remove or mitigate the caveats).</div><div class="gmail_extra"><br></div><div class="gmail_extra">- I want to have a "linker server". Usually this is accompanied by a significant amount of hand-waving w.r.t. the extent to which it could speeding up linking (or more generally, incremental builds). It's still not clear exactly how much speedup can be achieved by doing this, or even where the speedup will come from.</div><div class="gmail_extra">-- Some people focus a lot on the ability to do O(changes in the input) work. However, just like in the "libObjcopy" case, this is a much harder problem that it seems at first and for the same reason. And it's the same reason that "objcopy" is called "objcopy": really the only way to edit an object file is to copy it, which intrinsically does O(the output) work. At the tail end of the linker, there is a phase where the linker must iterate everything it needs to put in the output object file and assign it an offset; the offset of all later things depends on the offset of all earlier things. Then, it needs to iterate every relocation and look up the offset of the relocation target to do the relocation (LLD/ELF, in it's fastest mode, spends about 60% of its time doing this). Doing this from first principles incrementally is very, very hard; AFAIK, all existing "incremental link" approaches are fundamentally based on patching the previous output object file (which, among other things, does not produce deterministic outputs and generally feels like a hack). There are also even harder cases to make incremental, such as when the results of symbol resolution change or even worse the contents of archives end up changing.</div><div class="gmail_extra">-- From a larger perspective, there is a spectrum from fully statically linked binaries and just doing fine-grained dynamic libraries and letting the dynamic linker resolve all relocations. The details and the tradeoffs are what matter, and they are the things that have been least analyzed. Talking about a linker server generically is not that useful.</div><div class="gmail_extra"><br></div><div id="m_-1640572931879209436m_2604153376224275460DWT37198" class="gmail_extra">- I want to share code between JIT stuff in LLVM and the static "linker as a library". This has been brought up in this thread, if you look at the details, the argument is not as strong as the high-level software engineering spidey sense might initially suggest.</div></div></div></div></blockquote>Alright, let's be specific. PPC64TargetInfo::relocateOne in tools/lld/ELF/Target.cpp has this:<br><br> switch (Type) {<br> case R_PPC64_ADDR14: {<br> checkAlignment<4>(Val, Type);<br> // Preserve the AA/LK bits in the branch instruction<br> uint8_t AALK = Loc[3];<br> write16be(Loc + 2, (AALK & 3) | (Val & 0xfffc));<br> break;<br> }<br> case R_PPC64_ADDR16:<br> checkInt<16>(Val, Type);<br> write16be(Loc, Val);<br> break;<br> case R_PPC64_ADDR16_DS:<br> checkInt<16>(Val, Type);<br> write16be(Loc, (read16be(Loc) & 3) | (Val & ~3));<br> break;<br> ... (and so on; there are obviously lots of relocation types)<br><br>void RuntimeDyldELF::resolvePPC64Re<wbr>location in lib/ExecutionEngine/RuntimeDyl<wbr>d/RuntimeDyldELF.cpp has this:<br><br> switch (Type) {<br> case ELF::R_PPC64_ADDR14: {<br> assert(((Value + Addend) & 3) == 0);<br> // Preserve the AA/LK bits in the branch instruction<br> uint8_t aalk = *(LocalAddress + 3);<br> writeInt16BE(LocalAddress + 2, (aalk & 3) | ((Value + Addend) & 0xfffc));<br> } break;<br> case ELF::R_PPC64_ADDR16:<br> writeInt16BE(LocalAddress, applyPPClo(Value + Addend));<br> break;<br> case ELF::R_PPC64_ADDR16_DS:<br> writeInt16BE(LocalAddress, applyPPClo(Value + Addend) & ~3);<br> break;<br><br>and there are well over a hundred of these cases covering the different platforms we support. I don't feel like I need a special spider sense to tell me that this is a suboptimal bit of logic duplication. My primary reuse objective is to get rid of this duplicated logic for how to process relocations. There might be other opportunities as well (e.g. the logic to create stub functions and thunks).</div></div></blockquote><div><br></div><div>That seems bad code duplication. Maybe we should move this code to libObject and use it from both? One concern is error handling; the JIT linker handles broken relocations with assert() because that's an internal error, but in the static linker we want to always catch that to report to the user using our own error report mechanism.</div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div><div style="font-family:arial,helvetica,sans-serif;font-size:10pt;color:#000000"><span class="m_-1640572931879209436HOEnZb"><font color="#888888"><br><br> -Hal<br></font></span><blockquote style="border-left:2px solid rgb(16,16,255);margin-left:5px;padding-left:5px;color:rgb(0,0,0);font-weight:normal;font-style:normal;text-decoration:none;font-family:Helvetica,Arial,sans-serif;font-size:12pt"><span><div dir="ltr"><div class="gmail_extra"></div><div class="gmail_extra"><br></div><div class="gmail_extra"><div class="gmail_extra">- I actually don't want a linker in a library but I don't know enough about linkers to know that. What I actually want is "libObjcopy". I.e. some way to symbolically read, edit, and write object files. This is actually a much more complicated problem than the people asking for it think it is, and is unlikely to be as useful as they think it is. (it is still useful, but not as much as it might seem at first)</div></div><div class="gmail_extra"><br></div><div class="gmail_extra">- I want to have a COFF/ELF/MachO linker that shares a bunch of core functionality, such that each different format is just a "frontend" to the core linker, and a corresponding "backend".</div><div class="gmail_extra">-- From a practical perspective, the "atom LLD" tried this, and failed for a variety of reasons, one technical reason being that the ELF and COFF notion of "section" is a strict superset of its core "atom" abstraction (an indivisible chunk of data with a 1:1 correspondence with a symbol name). Therefore the entire design was impossible to use for those formats. In ELF and COFF, a section is decoupled from symbols, and there can be arbitrarily many symbols pointing to arbitrary parts of a section (whereas in MachO, "section" means something different; the analogous concept to an ELF/COFF section is basically "the thing that you get when you split a MachO section on symbol boundaries, assuming that the compiler/assembler hasn't relaxed relocations across those boundaries" which is not as powerful as ELF/COFF's notion). This resulted in severe contortions that ultimately made it untenable to keep working in that codebase. In an ideal world, the abstraction could have been fixed, but for various non-technical reasons that ended up not happening (I will not go into more detail about that on the mailing list). Also, interestingly, if you search the mailing list for "altentry", it seems like MachO has added a feature that breaks with the "atom" model as well. So there is some hope that further development on the atom LLD for MachO may fix this. Still, there are other challenges.</div><div class="gmail_extra"><br></div><div class="gmail_extra"><br></div><div class="gmail_extra">The experience from LLD/COFF and LLD/ELF means that we now have enough linker expertise in the LLVM community to think about these use cases seriously. However, this also means that we have enough linker expertise to drill down below the hand-waving, and often this means finding out that something that seemed reasonable or obvious is actually does not make as much sense as it seemed. It's important to not get upset when this happens.</div><div class="gmail_extra"><br></div><div class="gmail_extra"><br></div><div class="gmail_extra">This is totally orthogonal the discussion of whether ELF should be "llvm-like" in its modularity/library-like design, etc. If it was just me writing the linker, I would have followed a more traditional llvm-like design, but I am strongly convinced now that what was done was the correct decision at the time. If we had not mirrored the COFF design, Rui would not have felt as invested and we would have missed his contributions (I don't think that Rui has ever officially been working on ELF; what you see from him is true passion for the program). If we had designed ELF with a heavyweight traditional llvm library approach, we would have missed out of Rafael's contributions. Without these two there would not have been the critical mass needed to get an LLVM ELF linker off the ground and onto a trajectory leading to production. In its years of history, the biggest issue with getting LLD off the ground for ELF is that there was never a critical mass actively working on it, so getting this critical mass was actually quite important. Obviously there have been many others contributing. Especially Michael; without his expertise there was no way we would have gotten LLD working for PS4, which would have pulled Rafael, Michael, Davide, George Rimar, and myself (at the time) off of LLD. This would have meant that LLD never started showing enough signs of life for ELF and gotten FreeBSD onboard, which is critical to validating LLD as a production linker in open-source.</div><div class="gmail_extra"><br></div><div class="gmail_extra">The reality is that LLD/ELF is here, and its written in the way that it is written, and it got here the way that it got here. Patches are very welcome. Users with real use cases are very welcome. Technical discussion and evaluation of potential use cases is very welcome.</div><div class="gmail_extra"><br></div><div class="gmail_extra">-- Sean Silva</div></div>
<br></span><span>______________________________<wbr>_________________<br>LLVM Developers mailing list<br><a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br><a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br></span></blockquote><br><br><span><br>-- <br><div><span name="x"></span>Hal Finkel<br>Lead, Compiler Technology and Programming Languages<br>Leadership Computing Facility<br>Argonne National Laboratory<span name="x"></span><br></div></span></div></div></blockquote></div><br></div></div>