<br><br><div class="gmail_quote">On Tue Dec 09 2014 at 1:15:57 PM Akira Hatanaka <<a href="mailto:ahatanak@gmail.com">ahatanak@gmail.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 9, 2014 at 12:38 PM, Akira Hatanaka <span dir="ltr"><<a href="mailto:ahatanak@gmail.com" target="_blank">ahatanak@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><span>On Fri, Dec 5, 2014 at 2:40 PM, Eric Christopher <span dir="ltr"><<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><br><br><div class="gmail_quote"><span>On Wed Dec 03 2014 at 11:39:23 AM Akira Hatanaka <<a href="mailto:ahatanak@gmail.com" target="_blank">ahatanak@gmail.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 2, 2014 at 4:38 PM, Eric Christopher <span dir="ltr"><<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><br><br><div class="gmail_quote"><span>On Tue Dec 02 2014 at 4:31:43 PM Akira Hatanaka <<a href="mailto:ahatanak@gmail.com" target="_blank">ahatanak@gmail.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Tue, Dec 2, 2014 at 3:21 PM, Eric Christopher <span dir="ltr"><<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><span><div class="gmail_quote">On Mon Dec 01 2014 at 4:22:15 PM Bob Wilson <<a href="mailto:bob.wilson@apple.com" target="_blank">bob.wilson@apple.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word">Thanks for your feedback, Eric.<div><br></div><div>I still think we may be talking past each other a little bit, but rather than </div></div></blockquote><div><br></div></div></span><div class="gmail_quote"><div>Might be, sorry if so :(</div></div><div class="gmail_quote"><span><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div>delving further into the details right now, I’ve suggested to Akira that he look into how we should handle other kinds of options. I’m hoping that as we look at more of them, we will gain some insight into the approach that we want to take here. This patch really only deals with the easy cases, and whether we use Akira’s approach or not, I’m not especially worried about this kind of option.</div><div><br></div><div>Let’s come back to this in a little while after we’ve looked at what to do for some other options…..</div></div><div style="word-wrap:break-word"><div><br></div></div></blockquote><div><br></div></span><div>I apologize if this seems a bit rambly, but I'm trying to get a lot of thoughts on this down:</div><div><br></div></div><div class="gmail_quote"><div>FWIW I'm looking at some of this at the moment which is why I spoke to a few of the different options and needing to control TargetMachine creation. Just to continue with some of the discussion (at least to point some things the right direction and get feedback) let's think about the -target-abi option for ARM. It sets the ABI for the backend on the target machine. It probably shouldn't change on a per function basis (wait, except for the calling convention which can be changed via an attribute, fun!), so we want it to control global data layout and valid types which means we need it on the TargetMachine. The question is how to pass this information into the backend for such an option - right now, if we wanted, we could punt to using the command line options that currently exist and the existing ParseCommandLineOptions call from clang.</div><div><br></div></div></blockquote><div><br></div>
</div></div></div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><span>When option -target-abi=</span>apcs-gnu<span> for ARM is used for clang, </span>EmitAssemblyHelper::CreateTargetMachine passes FeatruresStr="+apcs" to Target::createTargetMachine. Do we have anything else we have to pass to the backend in this case?</div></div></div></div></blockquote><div><br></div></span><div>Not yet, but the option can be a backend option rather than a subtarget feature - and I'm probably going to do that in the short term as it's no worse than the current status, see the next line here below :)</div><span><font color="#888888"><div><br></div><div>-eric</div></font></span><div><div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">
<div> <br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div class="gmail_quote"><div></div><div>That said, it's a pretty awful hack to use them. For some of my particular needs I've got some code around unifying the DataLayout creation between the front end and back end, but it's a partial solution for the general problem and I'll probably run into more issues in this arena.</div><div><br></div><div>So, trying to come up with ways to communicate all of this via the API so that we can set ABIs for targets that need it (ARM, MIPS, others), as well as probably other options (anything passed via -backend-option is a good candidate).</div><div><br></div><div>One thought is expanding and/or deriving target specific versions of TargetOptions and populating that on creation from the front end. We'd want to be careful with the target specific bits and constructing defaults, probably something off of the bits in Targets.cpp would be appropriate. A lot of this type of code leads us down the TargetSupport library or something of the sort - classes to help describe or create backend constructs.</div><div><br></div><div>Anyhow, this is my current thinking on how to do API building of TargetMachine etc. (Alternately a TargetMachineBuilder? But that sounds complicated in all of the same ways without any particular gain).</div><span><font color="#888888"><div><br></div></font></span></div></blockquote></div></div></div></blockquote></div></div></div></blockquote><div><br></div><div><br></div></div></div></div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div>If we wanted to, in EmitAssemblyHelper::CreateTargetMachine, we could pass module-level CodeGenOpts.BackendOptions (backend options that don't need to change on a per-function basis) to Target::createTargetMachine. We can either pass it directly to the function or add a new field in llvm::TargetOptions that holds the options. If we can do that, we don't have to call ParseCommandLineOptions from EmitAssemblyHelper::CreateTargetMachine.</div></div></div></div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><br></div></div></div></div></blockquote><div><br></div></span><div>Yeah, I've been debating something like that. I'm not a huge fan of strings as a method of passing options around, but I'm starting to get worn down for sure. As far as ParseCommandLineOptions, the only deal is either we'd need to invent a new option format or we'd basically be calling it in the constructor for TargetMachine. I think I'd rather the new option format than that though. :\</div><div><br></div></div></blockquote><div><br></div></span><div>If we make changes in clang to pass the options to TargetMachine (via createTargetMachine) or save them in the IR, we can eventually remove the calls to ParseCommandLineOptions in clang. Is that right?</div><span><div><br></div></span></div></div></div></blockquote></div></div></div></blockquote><div><br></div><div>Ideally. I very much don't want to save options in the IR for more than functions as I mentioned before. We do it in a few cases now and it's pretty awkward.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><span><div></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="gmail_quote"><div></div><div>For my current problem I am going to end up just passing ABI as a string in TargetOptions, it's honestly the cleanest way to do it at the moment and it can be shared by multiple targets.</div><span><font color="#888888"><div><br></div></font></span></div></blockquote><div><br></div></span><div>I guess you can pass any other module-level options that you've mentioned that are passed via -backend-option in a similar fashion, perhaps using a key-value pair of strings. Is that what you have in mind?</div><div> </div></div></div></div></blockquote></div></div></div></blockquote><div><br></div><div>Basically. I'm unhappy with this solution, but a better one isn't happening at the moment.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div>Just to give you an update, I've been working on new patches based on the feedback I've received so far and discussing how I should proceed from here with Bob and Duncan. As the first step, we are trying to classify the command line options listed in the attachment to PR21471 into categories based on whether they are module-level or function level attributes or whether they affect subtarget creation, etc. I'll probably start new threads on llvm-dev soon to discuss the option categories individually to have a more focused discussion.</div><div><br></div></div></div></div></blockquote></div></div></div></blockquote><div><br></div><div>Sure. I'm interested in what you're thinking.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div></div><div>A couple of other questions:</div><div><br></div><div>- Subtarget lookup in XXXTargetMachine::getSubtargetImpl(const Function &F) seems a bit error-prone to me. If you add an option that is needed for subtargert creation as an attribute to Function, but forget to add it to the key that is used to search SubtargetMap, it can incorrectly return a cached subtarget when it should be creating a new one. Would it be better to generate a string key from the whole AttributeSet of the function (along with CPU and FeaturesStr and other options in TargetOptions, if necessary) and use it to do the lookup? The downside of course is that the key can get long as attributes that are not needed for subtarget creation are used to generate the key.</div><div><br></div></div></div></div></blockquote></div></div></div></blockquote><div><br></div><div>Yes, this is why I've only been enabling it for the subtarget dependent attributes. Also that way we don't create additional subtargets when we don't need to in cases where it doesn't matter.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div></div><div>- My understanding is that the only target currently using per-function subtarget creation is mips16/32. If that's correct, is there a reason you haven't enabled it for other targets, perhaps because you ran into or saw some problems?</div><div><div><div><br></div></div></div></div></div></div></blockquote></div></div></div></blockquote><div><br></div><div>It's enabled on more platforms than that now. Take a look at X86, aarch64, PPC...</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><div><div></div></div></div></div></div></div></blockquote><div><br></div></div></div></div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div>I see derived classes of TargetTransformInfo are calling TargetMachine::getSubtarget(Function*) to get the per-function subtarget in a few places, but other than those cases, TargetMachine's subtarget is used. </div></div></div></div></blockquote><div><br></div><div>Yep. It's a still in progress transition. The ARM abi bits are what's keeping me away at the moment, need to do that to move DataLayout back to TargetMachine so that I can use that in the middle end to do data layout.</div><div><br></div><div>-eric</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div> <br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><div><div></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="gmail_quote"><span><font color="#888888"><div></div><div>-eric</div></font></span><div><div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div></div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div class="gmail_quote"><div><div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div class="gmail_quote"><span><font color="#888888"><div></div><div>-eric</div></font></span></div><div><div><div class="gmail_quote"><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><div><blockquote type="cite"><div>On Dec 1, 2014, at 10:53 AM, Eric Christopher <<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.com</a>> wrote:</div><br><div>Hi Bob,<div><br></div><div>I've been giving this some thought, comments inline:<br><br><div class="gmail_quote">On Wed Nov 19 2014 at 6:24:18 PM Bob Wilson <<a href="mailto:bob.wilson@apple.com" target="_blank">bob.wilson@apple.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><blockquote type="cite"><div>On Nov 19, 2014, at 4:52 PM, Eric Christopher <<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.com</a>> wrote:</div><br><div><br style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;line-height:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><br style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;line-height:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><div class="gmail_quote" style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;line-height:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px">On Wed Nov 19 2014 at 4:39:42 PM Akira Hatanaka <<a href="mailto:ahatanak@gmail.com" target="_blank">ahatanak@gmail.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Wed, Nov 19, 2014 at 3:28 PM, Eric Christopher<span> </span><span dir="ltr"><<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.<u></u>com</a>></span><span> </span>wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div>So my general concern here is that lots of command line options that don't need to be or shouldn't be IR level constructs become oddly named string options. It's bad enough that we're doing that with the target cpu and target feature string let alone the rest of it.</div><div><br></div><div>If we're talking about things that end up changing the subtarget used we definitely need to make this part of the lookup key - at which point we're going to want a unified interface to all of them. I haven't applied your patch to see what the IR changes look like here.</div><div><br></div></blockquote><div><br></div></div></div></div><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div>With the patches applied, the attributes clang embeds in the IR look like this:</div><div><br></div><div>$ clang -cc1 ... -mllvm -arm-long-calls -mllvm -arm-reserve-r9 -menable-no-nans<br></div><div><p><span>attributes #0 = { nounwind readonly ssp "NoNaNsFPMath" "arm-long-calls"="1" "arm-reserve-r9"="1" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-<u></u>leaf" "no-infs-fp-math"="false" "stack-protector-buffer-size"=<u></u>"8" "target-cpu"="arm7tdmi" "target-features"="-neon,-<u></u>crypto,+apcs" "unsafe-fp-math"="false" "use-soft-float"="true" }</span></p><p>I'm guessing you want the front-end to convert command line options like "arm-long-calls" or "arm-reserve-r9" to something that doesn't look as odd?</p></div></div></div></div></blockquote><div><br></div><div>Right, that was part of the idea, the rest is that we define things that actually make sense as front end command line options and use those to communicate via IR constructs or api calls I.e. reserved registers don't need to be a specific back end option, but rather can be passed along when creating a target machine, etc.</div><div><br></div><div>-eric</div></div></div></blockquote><div><br></div></div></div><div style="word-wrap:break-word"><div><div>No, reserved registers (I assume you’re referring to “arm-reserve-r9” in Akira’s example) don’t necessarily need to be backend options. But, those kinds of options do need to be recorded in the IR and the target machine needs to reflect the relevant IR settings. Otherwise, we’ll still be left with a broken LTO model. E.G., I build one LTO object with “arm-reserve-r9” and another without it, but the LTO compilation still needs to respect those options for the code in each object. The details of how a particular option should be handled will vary depending on the option, and we need to discuss each one separately.</div><div><br></div></div></div></blockquote><div><br></div><div>ABI breaking options should probably be considered part of TargetMachine/Subtarget creation and handled accordingly (if it's subtarget creation it can be on the function as a feature string attribute, if it affects global code generation it needs to be at TargetMachine creation time).</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><div></div><div>It’s pretty clear, though, that there is a class of backend options that translate nicely to function attributes. This RFC is about the mechanism that we use to handle those (without specifying which specific options will be handled that way).</div></div></div><div style="word-wrap:break-word"><div><div><br></div></div></div></blockquote><div><br></div><div>Fair enough. For function attributes I can see a desire to have (as Chris posted in an email from someone at synopsys) a generic target specific dictionary mapping on functions for backend specific uses. I just haven't seen anything in the thread that would be applicable for this rather than an actual function attribute. That said, I would like some sort of delineation for target function attributes as well. Thoughts on that? Perhaps it could look like the string dictionary we've mentioned, maybe not.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><div></div><blockquote type="cite"><div><div class="gmail_quote" style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;line-height:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div></div><div>From looking at the patch it seems that the command line option for NaNs should be a general attribute on the function rather than a TargetMachine option anyhow, so it seems that should be migrated down. Whether or not it is exposed as an llvm attribute or a string based key value pair is an interesting discussion. Also, would we want to, instead, expose the backend options as clang IR level options? Seems to make sense from that perspective and then they can add the appropriate attributes to the functions themselves. (-mlongcalls for the long calls option for example).</div><div><br></div><div>Thanks!</div><span><font color="#888888"><div><br></div><div>-eric</div></font></span></blockquote></div></div></div></blockquote></div></div></blockquote><div><br></div></div></div><div style="word-wrap:break-word"><div>Again, I think we can evaluate specific options on a case-by-case basis. For some of them, it might make sense to expose them as clang options. Like you, I had previously assumed we would teach the front-end about all the backend options. I was a little surprised by the approach Akira suggested here, but after giving it some thought, I think it’s a great idea. There are a LOT of target-specific backend options. Some of them do not need to be IR, but many do. Teaching clang about all of them would be a lot of work, and I don’t see any clear benefit from that. This approach lets clang remain relatively independent of the backend options. We still need to think carefully about which backend options use this mechanism, of course.</div><div><br></div></div></blockquote><div><br></div><div>I'm unconvinced here...</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div></div><div>Are there other reasons why you think we should add all those options to clang? I just don’t see much to be gained by that.</div><div><br></div></div></blockquote><div><br></div><div>.... I think that anything you'd like to expose should probably be a front end option. If nothing else I don't see how you plan on annotating the IR. Basically you're constructing a parallel option hierarchy and that seems... redundant. I agree that it's nice for tools to be able to override things in the backend - for testing if nothing else. We do this a lot by checking options and values in the code. A way of streamlining this would be nice - but I don't think that's the goal of these patches no?</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div></div><div>Do you have objections to the use of string-based key/value pairs? I believe the intention all along, going back to Bill’s work on this a few years ago, was that we would add new attributes for target-independent things and use strings for the target-specific options. That seems like a reasonable starting point to me.</div></div><div style="word-wrap:break-word"><div><br></div></div></blockquote><div><br></div><div>Only some as I mentioned above - I haven't seen anything that needs to be target specific in this way, but I'm open to the idea :)</div><div><br></div><div>As a note, this is all for function level things. We need a solution for non-function/module level attributes and I think that's going to look like API calls into the backend for TargetMachine creation - there may be a list of things that would turn into module level attributes, but that is a direction fraught with danger and I'd like to avoid that as much as possible.</div><div><br></div><div>(As a note, the particular patch that Akira has is unacceptable because of the changes to the subtarget lookup, but that's a separate issue :)</div><div><br></div><div>-eric</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div style="word-wrap:break-word"><div><blockquote type="cite"><div><div class="gmail_quote" style="font-family:Helvetica;font-size:12px;font-style:normal;font-variant:normal;font-weight:normal;letter-spacing:normal;line-height:normal;text-align:start;text-indent:0px;text-transform:none;white-space:normal;word-spacing:0px"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div><div><br><div class="gmail_quote">On Tue Nov 18 2014 at 12:26:49 PM Akira Hatanaka <<a href="mailto:ahatanak@gmail.com" target="_blank">ahatanak@gmail.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><div dir="ltr"><div><div>Updated patch is attached. Note this is just a work-in-progress patch and I plan to address the feedback comments later if this patch is in the right direction.<br></div><div><br></div><div>This is how the command line options are parsed and used by the backend passes:</div><div><br></div><div>1. Tools such as clang and llc call cl::ParseCommandLineOptions. Any of the options that should be written to the bitcode as function or module attributes (these options are declared as CodeGenOpt, which is a subclass of cl::opt) are saved to Function or ModuleIntAttrs in CodeGenOptions.cpp.</div><div>2. setFunctionAttribute is called to move the function and module options saved in step 1 to the bitcode. Also, options stored in TargetOptions are moved to the bitcode too. Pre-existing attributes in the bitcode are overwritten by the options in the command line. </div><div>3. The backend passes and subtarget classes call getFunctionAttribute to read the attributes stored to the bitcode in step 2. Function::hasFnAttribute can be called directly (for example, NoNaNsFPMath in the patch), if the default value is known to be "false".</div><div><br></div><div>I also made the following changes in the patch:</div><div><br></div><div>1. In the constructor of MachineFunction, call the version of getSubtargetImpl that takes a Function parameter, so that it gets the Subtarget specific to the Function being compiled.</div><div>2. Change ARMTargetMachine::SubtargetMap to be a DenseMap<Function*, unique_ptr<ARMSubtarget>>. This is just a temporary change to ease debugging and should be reverted to a StringMap or changed to another type of map later.</div><div><br></div></div></div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Nov 17, 2014 at 11:40 AM, Eric Christopher<span> </span><span dir="ltr"><<a href="mailto:echristo@gmail.com" target="_blank">echristo@gmail.<u></u>com</a>></span><span> </span>wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><br><br><div class="gmail_quote"><span>On Mon Nov 17 2014 at 9:56:40 AM Bob Wilson <<a href="mailto:bob.wilson@apple.com" target="_blank">bob.wilson@apple.com</a>> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><br>> On Nov 14, 2014, at 2:44 PM, Duncan P. N. Exon Smith <<a href="mailto:dexonsmith@apple.com" target="_blank">dexonsmith@apple.com</a>> wrote:<br>><br>> +chrisb<br>><br>>> On 2014-Nov-13, at 16:33, Akira Hatanaka <<a href="mailto:ahatanak@gmail.com" target="_blank">ahatanak@gmail.com</a>> wrote:<br>>><br>>> I'm working on fixing PR21471, which is about embedding codegen command line options into the bitcode as function or module-level attributes so that they don't get ignored when doing LTO.<br>>><br>>><span> </span><a href="http://llvm.org/bugs/show_bug.cgi?id=21471" target="_blank">http://llvm.org/bugs/show_<u></u>bug.<u></u>cgi?id=21471</a><br>>><br>>> I have an initial patch (attached to this email) which enables clang/llvm to recognize one command line option, write it to the IR, and read it out in a backend pass. I'm looking to get feedback from the community on whether I'm headed in the right direction or whether there are alternate ideas before I go all the way on fixing the PR. Specifically, I'd like to know the answers to the following questions:<br>>><br>>> 1. How do we make sure we continue to be able to use the command line options we've been using for llc and other tools?<br>>> 2. How to handle cases where two functions in a module have different sets of command line options?<br>>> 3. Where should the command line options or module/function attributes be stored once they are read out from the IR?<br>>><br>>> The short description of the approach I took in my patch is that command line options that are important to codegen are collected by cl::ParseCommandLineOptions, written to the bitcode as function or module attributes, and read out directly by the optimization passes that need them. cl::opt options are replaced with CodeGenOpt options (subclass of cl::opt), which are needed only to parse the command line and provide the default value when the corresponding options are not in the bitcode.<br>><br>> I like this approach, since it means the frontend doesn't have to understand<br>> options in order to pass them on to the backend.<br><br>I agree. It’s not the approach that was I was expecting, but after reading the patches, I like it.<br><br>If we can get consensus on the approach, I suggest that we stage the work to first adopt Akira’s patches and then we can incrementally convert various options over to use it. There may be complications for some options, so I’d like to see separate patches for each one (possibly with a few similar options combined in one patch).<br><br></blockquote><div><br></div></span><div>I'm not sold quite yet :)</div><div><br></div><div>Akira: Can you show an example of how you expect this to work in the IR and how the backend is going to process?</div><div><br></div><div>Thanks.</div><span><font color="#888888"><div><br></div><div>-eric</div></font></span><div><div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">><br>> The static variables should be straightforward to migrate to an LLVMContext<br>> once ParseCommandLineOptions stores things there instead of in globals.<br>><br>>> diff --git a/lib/CodeGen/CodeGenOption.<u></u>cp<u></u>p b/lib/CodeGen/CodeGenOption.<u></u>cp<u></u>p<br>>> new file mode 100644<br>>> index 0000000..2d74c2f<br>>> --- /dev/null<br>>> +++ b/lib/CodeGen/CodeGenOption.<u></u>cp<u></u>p<br>>> @@ -0,0 +1,59 @@<br>>> +//===- CodeGen/CodeGenOptions.cpp - Code-gen option. --*- C++ -*-===//<br>>> +//<br>>> +// The LLVM Compiler Infrastructure<br>>> +//<br>>> +// This file is distributed under the University of Illinois Open Source<br>>> +// License. See LICENSE.TXT for details.<br>>> +//<br>>> +//===------------------------<u></u><u></u>------------------------------<u></u><u></u>----------------===//<br>>> +//<br>>> +//===------------------------<u></u><u></u>------------------------------<u></u><u></u>----------------===//<br>>> +<br>>> +#include "llvm/CodeGen/CodeGenOption.h"<br>>> +#include "llvm/IR/Attributes.h"<br>>> +#include "llvm/IR/Module.h"<br>>> +<br>>> +using namespace llvm;<br>>> +<br>>> +static std::map<std::string, bool> FunctionBoolAttrs;<br>>> +static std::map<std::string, bool> ModuleBoolAttrs;<br>>> +<br>><br>> @Chris, should these be using ManagedStatic?<br>><br>> ______________________________<u></u><u></u>_________________<br>> LLVM Developers mailing list<br>><span> </span><a href="mailto:LLVMdev@cs.uiuc.edu" target="_blank">LLVMdev@cs.uiuc.edu</a> <a href="http://llvm.cs.uiuc.edu/" target="_blank">http://llvm.cs.uiuc.edu</a><br>><span> </span><a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev" target="_blank">http://lists.cs.uiuc.edu/<u></u>mai<u></u>lman/listinfo/llvmdev</a><br><br><br>______________________________<u></u><u></u>_________________<br>LLVM Developers mailing list<br><a href="mailto:LLVMdev@cs.uiuc.edu" target="_blank">LLVMdev@cs.uiuc.edu</a> <a href="http://llvm.cs.uiuc.edu/" target="_blank">http://llvm.cs.uiuc.edu</a><br><a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev" target="_blank">http://lists.cs.uiuc.edu/<u></u>mailm<u></u>an/listinfo/llvmdev</a><br></blockquote></div></div></div><br>______________________________<u></u>_________________<br>LLVM Developers mailing list<br><a href="mailto:LLVMdev@cs.uiuc.edu" target="_blank">LLVMdev@cs.uiuc.edu</a> <a href="http://llvm.cs.uiuc.edu/" target="_blank">http://llvm.cs.uiuc.edu</a><br><a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev" target="_blank">http://lists.cs.uiuc.edu/<u></u>mailman/listinfo/llvmdev</a></blockquote></div></div></blockquote></div></div></div></blockquote></div></div></div></blockquote></div></div></blockquote></div><br></div></blockquote></div></div>
</div></blockquote></div><br></div></div></blockquote></div>
</div></div></blockquote></div></div></div></blockquote></div></div></div>
</blockquote></div></div></div></blockquote></div></div></div>
</blockquote></div></div></div><br></div></div>
</blockquote></div></div></div></blockquote></div>