r230423 - Wrap clang module files in a Mach-O, ELF, or COFF container.

[Adrian Prantl]

> On Feb 25, 2015, at 3:37 PM, Richard Smith <richard at metafoo.co.uk> wrote:
> 
> On Wed, Feb 25, 2015 at 2:22 PM, Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
> 
> > On Feb 25, 2015, at 1:34 PM, Richard Smith <richard at metafoo.co.uk <mailto:richard at metafoo.co.uk>> wrote:
> >
> > On Wed, Feb 25, 2015 at 1:02 PM, Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
> > On Feb 24, 2015, at 6:28 PM, Adrian Prantl <aprantl at apple.com <mailto:aprantl at apple.com>> wrote:
> >> You are correct. I need to remove the use of PCHGenerator from ModuleContainerGenerator.
> >>> On Feb 24, 2015, at 6:11 PM, NAKAMURA Takumi <geek4civic at gmail.com <mailto:geek4civic at gmail.com>> wrote:
> >>> It still has circular dependencies between clangCodeGen and clangFrontend.
> >
> > After some deliberation, I noticed that there is actually more to be done here in order to resolve the cyclic dependencies.
> >
> > Well, it makes sense that there would be problems here, since you're coupling our lex-only / parse-only modes to IR generation. I wonder if it would be possible to entirely isolate Frontend from knowledge of the module wrapper format (putting it in FrontendTool instead)?
> 
> I was able to make PCHGenerator agnostic of the wrapper format, but it is not possible to push the wrapper-awareness out any further than FrontendActions.
> 
> "not possible" is quite a strong claim to make. It would seem possible for FrontendTool to hand the CompilerInstance or Action a collection of callbacks for creating / importing module file data.

Challenge accepted, I should be possible: I image that we could introduce a ModuleContainerGeneratorFactory (although I find this a little revolting ;-) that is passed to the CompilerInstance, which then hands it down to GenerateModuleAction/GeneratePCHAction, which asks the Factory to generate an new ModuleContainerGenerator instance which is actually an LLVModuleContainerGenerator. This way Frontend only needs to link against whereever the ModuleContainerGenerator base class is defined and each tool can decide which backend to use.

> 
> > If not, I don't think we have any other option than to grow a Frontend->CodeGen dependency, which in turn will be terrible for people who want to use our frontend with a different backend...
> 
> Is that a hypothetical scenario or are there any such users out there?
> 
> I don't have concrete knowledge of anyone doing this today, but enough people have asked about this that it's not unreasonable to think that someone might have actually done it.
>  
> I understand there are tools like clang-format/modernize/... which after this change have to link against the LLVM targets in order to generate clang-compatible modules; but are there any non-LLVM compilers that use clang as a frontend?
> 
> I think there are, but I'm not sure if they all do this by converting LLVM IR to their own IR or whether some of them go straight from a Clang AST. Whether or not such compilers exist today, we should try to avoid creating barriers for people who want to do this in the future.
> 
> > Here is a graph of (a simplified subset of) the dependencies after this commit:
> > - Pretty much all of CodeGen depends on CodeGenOptions, which is currently part of Frontend.
> > - BackendUtil and CodeGenAction depend on both CodeGen and Frontend.
> > - CodeGenModuleContainer introduces a cyclic dependency between Frontend and CodeGen.
> >
> > <before.png>
> >
> > The above cycle can be resolved by reversing the CodeGen->Frontend dependency and splitting out the common dependencies CodeGenOptions and frontend::utils::BuryPointer into a separate library that I’m calling FrontendSupport for lack of a better name.
> >
> > The right place for CodeGenOptions is probably Basic, alongside LangOptions, TargetOptions, CommentOptions, etc.
> 
> That sounds like a good idea; I might also be able to move it into CodeGen itself.
> >
> > After this, the only remaining CodeGen->Frontend dependencies are CodeGen/BackendUtil.cpp and CodeGen/CodeGenAction.cpp:
> > - CodeGenAction looks like it could safely be moved into FrontendTool, which is its only user.
> >
> > I don't think that's necessarily a good idea: CodeGenAction is tightly coupled to CodeGen and only very loosely coupled to the frontend.
> 
> I agree that it is tightly coupled to CodeGen, but it is also tightly coupled to CompilerInstance, which has its tentacles all over Frontend so it needs to move somewhere to break the cycle. Do you see any specific problems with having CodeGenAction in FrontendTool?
> 
> No specific problem, only a general malaise: we deliberately isolate all the parts of Clang that touch LLVM IR in lib/CodeGen; this change would violate that notional encapsulation.

Okay, separating out everything that touches IR that makes sense. I think it should be possible to remove all references to CompilerInstance from CodeGenAction; it appears to be only used as a holder of the various options and the DiagnosticsEngine, which would allow us to leave it in CodeGen.

I’ll try that.
-- adrian

> 
> thanks for the feedback!
> adrian
> >
> > - BackendUtil can stay were it is, it is needed by CodeGenAction and (via CodeGenModuleContainer) by Frontend. The dependency on Frontend can be eliminated by splitting BuryPointer out from Utils.
> > The new picture then looks like this:
> >
> > <after.png>
> >
> > I’ll try and implement it this way; hopefully I didn’t miss any other edges in the graph.
> > -- adrian
> >
> >>
> >> thanks for noticing!
> >> -- adrian
> >
> >
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20150225/b1b8337f/attachment.html>