llvm dev - Jun 2015 - [LLVMdev] Linking modules across contexts crashes

> On 2015-Jun-01, at 11:06, Reid Kleckner <rnk at google.com> wrote:
> 
> I'm pretty sure module linking is expected to occur in the same LLVM
context.
Correct.
> IIRC Duncan had some proposal for how ld64 could do something clever with
multiple contexts, but I've totally forgotten what it was.
This was for LTO (probably unrelated to Yuri's scenario?).

 1. Lazy-load each module in its own context, parse its symbols
    (without actually loading functions, metadata, etc.), and
    destroy the module and context.
 2. Decide which modules we should link (based on symbol parsing).
 3. Load all the chosen modules into a single context and link them
    together.

lib/LTO has a bunch of the pieces for this, but I got the idea from
Rafael's work on tools/gold-plugin/gold-plugin.cpp.  It might be
easier to look at there.
> On Fri, May 29, 2015 at 5:18 PM, Yuri <yuri at rawbw.com> wrote:
> I get a crash when I try to link multiple modules registered in their
individual contexts.
> Documentation for Linker::LinkModules doesn't mention anything about
contexts, and the first link succeeds. But the second link crashes.
The problem is that a fair number of things referenced by a Module
are owned by its LLVMContext (types, constants, metadata, and some
other things).
> Is this not the right way to merge such modules? If not, then what is the
right way?
You can round-trip to bitcode, reading the module into the
destination context.  The following pseudo-code gives the idea:

    bool linkModuleFromDifferentContext(Module &D, const Module &S) {
      SmallVector<char, 256> Buffer;
      writeBitcodeToBuffer(S, Buffer);

      std::unique_ptr<Module> M = readBitcodeFromBuffer(D.getContext());
      return Linker::LinkModules(&D, M.get());
    }
> In any case, documentation for Linker::LinkModules should say if contexts
are or aren't expected to be the same.
Good idea; patch welcome.
> 
> Yuri
> 
> 
> ---testcase---
> #include "llvm/IR/LLVMContext.h"
> #include "llvm/IR/Module.h"
> #include "llvm/Linker/Linker.h"
> #include "llvm/Support/raw_ostream.h"
> #include "llvm/IR/DiagnosticPrinter.h"
> #include <string>
> #include <iostream>
> 
> using namespace llvm;
> using namespace std;
> 
> int main() {
>   // vars
>   string Message;
>   raw_string_ostream Stream(Message);
>   DiagnosticPrinterRawOStream DP(Stream);
>   LLVMBool Result;
>   // create blank modules and contexts
>   LLVMContext *ctx1 = new LLVMContext;
>   Module* module1 = new Module("module1", *ctx1);
>   LLVMContext *ctx2 = new LLVMContext;
>   Module* module2 = new Module("module2", *ctx2);
>   LLVMContext *ctx3 = new LLVMContext;
>   Module* module3 = new Module("module3", *ctx3);
>   // fill modules
> llvm::Function::Create(llvm::FunctionType::get(Type::getInt32Ty(*ctx1),
false), llvm::Function::ExternalLinkage, "f1", module1);
> llvm::Function::Create(llvm::FunctionType::get(Type::getInt32Ty(*ctx2),
false), llvm::Function::ExternalLinkage, "f2", module2);
> llvm::Function::Create(llvm::FunctionType::get(Type::getInt32Ty(*ctx3),
false), llvm::Function::ExternalLinkage, "f3", module3);
> 
>   // merge f1 <- f2
>   Result = Linker::LinkModules(module1, module2, [&](const
DiagnosticInfo &DI) {DI.print(DP);});
>   cout << "merge result=" << Result << endl;
>   delete ctx2;
>   cout << "--done merge #1--" << endl;
> 
>   // merge f3 <- f1
>   Result = Linker::LinkModules(module3, module1, [&](const
DiagnosticInfo &DI) {DI.print(DP);});
>   cout << "merge result=" << Result << endl;
>   delete ctx1;
>   cout << "--done merge #2--" << endl;
> 
>   return 0;
> }
> 
> ---output---
> merge result=0
> --done merge #1--
> Bus error
> 
> rev.237344
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>

On 06/01/2015 11:43, Duncan P. N. Exon Smith wrote:
> You can round-trip to bitcode, reading the module into the
> destination context.  The following pseudo-code gives the idea:
>
>      bool linkModuleFromDifferentContext(Module &D, const Module
&S) {
>        SmallVector<char, 256> Buffer;
>        writeBitcodeToBuffer(S, Buffer);
>
>        std::unique_ptr<Module> M =
readBitcodeFromBuffer(D.getContext());
>        return Linker::LinkModules(&D, M.get());
>      }
Duncan,

Thanks for this workaround, i works. However, going to binary and back 
causes the significant bump in the process user time (14s->16s), while 
the wall clock time still lower due to parallelization.

There should be the function llvm::MergeContexts, similar to 
llvm::Linker::LinkModules. It should merge contexts as containers, with 
an error when this isn't possible, without changing most of the objects 
at all. It should cause the source context to disappear.

Yuri

> On 2015-Jun-02, at 12:37, Yuri <yuri at rawbw.com> wrote:
> 
> On 06/01/2015 11:43, Duncan P. N. Exon Smith wrote:
>> You can round-trip to bitcode, reading the module into the
>> destination context.  The following pseudo-code gives the idea:
>> 
>>     bool linkModuleFromDifferentContext(Module &D, const Module
&S) {
>>       SmallVector<char, 256> Buffer;
>>       writeBitcodeToBuffer(S, Buffer);
>> 
>>       std::unique_ptr<Module> M =
readBitcodeFromBuffer(D.getContext());
>>       return Linker::LinkModules(&D, M.get());
>>     }
> 
> Duncan,
> 
> Thanks for this workaround, i works. However, going to binary and back
causes the significant bump in the process user time (14s->16s), while the
wall clock time still lower due to parallelization.
> 
> There should be the function llvm::MergeContexts, similar to
llvm::Linker::LinkModules. It should merge contexts as containers, with an error
when this isn't possible, without changing most of the objects at all. It
should cause the source context to disappear.
If you want to work on a solution, I think a better approach would be
adding API for one (or both) of:

    class Module {
    public:
        /// Move this module to the given context.
        void changeContext(LLVMContext &NewContext);
    };

    /// Remap the given module in a different context.
    std::unique_ptr<Module> remapModuleToContext(
        const Module &M, LLVMContext &NewContext);

I'm not sure what your exact use case is, but I could see either/both
interfaces being useful.  For example, you could change
`Linker::LinkModules()` to something like:

    bool Linker::LinkModules(Module &Dest, Module &Src) {
      if (&Src.getContext() != &Dest.getContext())
        // Recreate Src in Dest's context and link that.
        return LinkModules(
                   Dest,
                   remapModuleToContext(Src, Dest.getContext()).get());

      // Old code...
    }

or:

    bool Linker::LinkModules(Module &Dest, Module &Src) {
      if (&Src.getContext() != &Dest.getContext())
        // Destructively move Src to Dest's context.
        Src.changeContext(Dest.getContext());

      // Old code...
    }

Patches welcome!