[llvm] r208257 - [RuntimeDyld] Make RuntimeDyldImpl::resolveExternalSymbols preserve the

Kaylor, Andrew andrew.kaylor at intel.com
Thu May 8 11:58:12 PDT 2014


Tiering of optimization levels is obviously a valid use case.  I think it would be possible to work support for that back into MCJIT, but with multiple module support relinking would be a bit of a nightmare.  Off the cuff, I'd say you'd need some kind of stub or offset table system to handle it.

Personally, I really like the way that WebKit handles that by not letting MCJIT control things it doesn't understand.

-Andy

From: Jim Grosbach [mailto:grosbach at apple.com]
Sent: Wednesday, May 07, 2014 5:38 PM
To: Kaylor, Andrew
Cc: Lang Hames; llvm-commits at cs.uiuc.edu
Subject: Re: [llvm] r208257 - [RuntimeDyld] Make RuntimeDyldImpl::resolveExternalSymbols preserve the

FWIW, support for recompilation of modules was a conscious intention of the design originally in order to support tiering up levels of optimization. The implementation since then has obviously diverged from that a bit, but it's still a relevant design goal.

You make a good point on the section permissions. Supporting recompilation is more than just relocation handling and would require nontrivial juggling in a memory manager to do effectively.

-Jim

On May 7, 2014, at 4:43 PM, Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>> wrote:


I'm not sure I'd agree that MCJIT is supposed to support recompilation of a module.  I believe we've explicitly said that you can't modify a module after it has been added to MCJIT, so why would you recompile it?  In any event, if you do recompile a module, the relocation information will be recreated.

But I take it that you're actually addressing some case where a module is moved after it has been compiled and put into an executable state.  I don't understand this case either.  Why would you move a module after it has been made executable?

Also, I would expect that you'll run into problems with memory managers that set the permissions for code sections to RX after a module has been finalized.  In those cases, if you try to reapply relocations (presumably to a module that didn't move) it will cause a protection fault.  Conversely, if you allow a module to be moved after it has been finalized, don't you run into issues with the client having to keep track of whether or not the module is in an executable state?

-Andy

From: Lang Hames [mailto:lhames at gmail.com]
Sent: Wednesday, May 07, 2014 4:09 PM
To: Kaylor, Andrew
Cc: llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>
Subject: Re: [llvm] r208257 - [RuntimeDyld] Make RuntimeDyldImpl::resolveExternalSymbols preserve the

Hi Andy,

MCJIT is supposed to support recompilation of a module, and linking multiple modules (thanks for making the latter work! :). So we need to keep external relocations around in case the module they point to is recompiled.

Cheers,
Lang.

On Wed, May 7, 2014 at 3:56 PM, Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>> wrote:
The discarding of processed relocations was an intentional choice.  The relocation map can take up a pretty fair amount of memory, particularly in the case where you have a large number of modules that reference one another.  Keeping the relocation map around will cause memory bloat issues for some clients.

My thinking was that once a module was finalized you can't remap its sections anymore.  Do you have a use case where this is an unacceptable limitation?

-Andy

-----Original Message-----
From: llvm-commits-bounces at cs.uiuc.edu<mailto:llvm-commits-bounces at cs.uiuc.edu> [mailto:llvm-commits-bounces at cs.uiuc.edu<mailto:llvm-commits-bounces at cs.uiuc.edu>] On Behalf Of Lang Hames
Sent: Wednesday, May 07, 2014 3:34 PM
To: llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>
Subject: [llvm] r208257 - [RuntimeDyld] Make RuntimeDyldImpl::resolveExternalSymbols preserve the

Author: lhames
Date: Wed May  7 17:34:08 2014
New Revision: 208257

URL: http://llvm.org/viewvc/llvm-project?rev=208257&view=rev
Log:
[RuntimeDyld] Make RuntimeDyldImpl::resolveExternalSymbols preserve the relocation entries it applies.

Prior to this patch, RuntimeDyldImpl::resolveExternalSymbols discarded relocations for external symbols once they had been applied. This causes issues if the client calls MCJIT::finalizeLoadedModules more than once, and updates the location of any symbols in between (e.g. by calling MCJIT::mapSectionAddress).

No test case yet: None of our in-tree memory managers support moving sections around. I'll have to hack up a dummy memory manager before I can write a unit test.

Fixes <rdar://problem/16764378>


Modified:
    llvm/trunk/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp

Modified: llvm/trunk/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp?rev=208257&r1=208256&r2=208257&view=diff
==============================================================================
--- llvm/trunk/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp (original)
+++ llvm/trunk/lib/ExecutionEngine/RuntimeDyld/RuntimeDyld.cpp Wed May
+++ 7 17:34:08 2014
@@ -620,6 +620,8 @@ void RuntimeDyldImpl::resolveRelocationL
 }

 void RuntimeDyldImpl::resolveExternalSymbols() {
+  StringMap<RelocationList> ProcessedSymbols;
+
   while (!ExternalSymbolRelocations.empty()) {
     StringMap<RelocationList>::iterator i = ExternalSymbolRelocations.begin();

@@ -665,8 +667,20 @@ void RuntimeDyldImpl::resolveExternalSym
       resolveRelocationList(Relocs, Addr);
     }

+    ProcessedSymbols[i->first()] = i->second;
     ExternalSymbolRelocations.erase(i);
   }
+
+  // Restore the relocation entries that were consumed in the loop above:
+  //
+  // FIXME: Replace the following loop with:
+  //           std::swap(ProcessedSymbols, ExternalSymbolRelocations)
+  //        once StringMap has copy and move construction.
+  for (StringMap<RelocationList>::iterator I = ProcessedSymbols.begin(),
+                                           E = ProcessedSymbols.end();
+       I != E; ++I) {
+    ExternalSymbolRelocations[I->first()] = I->second;  }
 }
 //===----------------------------------------------------------------------===//


_______________________________________________
llvm-commits mailing list
llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits

_______________________________________________
llvm-commits mailing list
llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>
http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140508/a1ebba15/attachment.html>


More information about the llvm-commits mailing list