[llvm-dev] Linking existing functions from JITed code

[Lang Hames via llvm-dev]

Hi Andy,

Rather than adding these to the object linking layer I'd be inclined to add
a new global-mapping layer that handles the global mappings and nothing
else. Clients could then compose this with the linking layer if they want
the global mapping functionality.

Any interest in writing up a patch for something like this? If so I'd be
happy to review it, otherwise I'll probably end up writing something along
these lines during the Kaleidoscope cleanup process.

Cheers,
Lang.


On Wed, Aug 19, 2015 at 8:23 PM, Andy Somogyi <andy.somogyi at gmail.com>
wrote:

> Lang,
>
> I added the add/get global mapping to my kaleidoscope JIT, but I think
> perhaps these would make more sense if they were added to the object
> linking layer as they would be generally usable there.
>
>
> On Aug 19, 2015, at 11:19 PM, Andy Somogyi wrote:
>
> Hey Lang,
>
> I've added this to my Kaleidoscope JIT, and it seems to work just fine,
> basically I copied the global mapping functionality out of execution engine
> and hooked up the lambda resolver to reference them.
>
> Here's the code if you want it:
>
> https://github.com/AndySomogyi/cayman/blob/master/src/llvm_orc_initial.cpp
>
>
> On Aug 19, 2015, at 9:14 PM, Lang Hames wrote:
>
> Hi Andy,
>
> I think that makes sense. I'm currently rewriting the core Kaleidoscope
> tutorials - I'll look at adding support for this.
>
> - Lang.
>
> On Fri, Aug 14, 2015 at 7:38 AM, Andy Somogyi <andy.somogyi at gmail.com>
> wrote:
>
>> After some fiddling with it, it does in fact look like it works as you
>> describe Lang.
>>
>> The trick was you had to call
>>
>> llvm::sys::DynamicLibrary::LoadLibraryPermanently(nullptr);
>>
>> to add the currently running process before calling
>>
>> llvm::RTDyldMemoryManager::getSymbolAddressInProcess(name);.
>>
>> Also, all of the function needs to be declared as extern C to avoid name
>> mangling.
>>
>> I however think that adding the
>> addGlobalMapping/getPointerToGlobalMapping functions to the ORC
>> kaleidoscope examples is very useful, as for example one wants a set of
>> private symbols made available to JITed code.
>>
>>
>>
>> On Aug 13, 2015, at 7:41 PM, Lang Hames <lhames at gmail.com> wrote:
>>
>> Hi Andy,
>>
>> I haven't tested this on Linux, but on MacOS the
>> RuntimeDyldMemorManager::getSymbolAddressInProcess method should find
>> symbol addresses in the host program, including symbols from static
>> archives linked into the program. However, one gotcha is that the symbol
>> has to be reachable from main, otherwise the linker may strip it from the
>> final executable.
>>
>> Do you have a test-case that I could try to reproduce the issue with?
>>
>> Alternatively, if you run the code under a debugger, do you see symbols
>> that the JIT failed to find? If the symbols are visible in the debugger but
>> invisible to the JIT that sounds like a bug in the JIT. If the symbols are
>> invisible in both that would suggest that the linker is stripping them out.
>>
>> - Lang.
>>
>>
>> On Thu, Aug 13, 2015 at 1:12 PM, David Blaikie <dblaikie at gmail.com>
>> wrote:
>>
>>> +"Lang Hames, Linker of Linkers"
>>>
>>> (Lang developed the ORC JIT you seem to be using & can provide more
>>> context than I have)
>>>
>>> On Thu, Aug 13, 2015 at 12:43 PM, Andy Somogyi via llvm-dev <
>>> llvm-dev at lists.llvm.org> wrote:
>>>
>>>> Hi
>>>>
>>>> I’ve previously used the ExecutionEngine::addGlobalMapping to make
>>>> existing functions available to my JITed code.
>>>>
>>>> I’m currently using ORC, as MCJIT does not appear to be maintained any
>>>> longer (the kaleidoscope examples have not worked for some time with
>>>> MCJIT).
>>>>
>>>> I’m using just the basic ORC CompileLayer directly.
>>>>
>>>> So, I’ve essentially copied the ExecutionEngine::addGlobalMapping
>>>> related function to my JIT context, and I create a lambda resolver as such:
>>>>
>>>> JITContext::addModule(…) {
>>>>
>>>> auto Resolver = createLambdaResolver(
>>>> [&](const std::string &name) {
>>>>
>>>> // look up first in JIT'ed code
>>>> if (auto sym = findMangledSymbol(name)) {
>>>> return RuntimeDyld::SymbolInfo(sym.getAddress(),
>>>> sym.getFlags());
>>>> return RuntimeDyld::SymbolInfo(nullptr);
>>>> }
>>>>
>>>> // look up in added globals
>>>> if (auto addr = getPointerToGlobalMapping(name)) {
>>>> return RuntimeDyld::SymbolInfo(addr, JITSymbolFlags::Exported);
>>>> }
>>>>
>>>> // finally try to look up existing process symbols, note
>>>> // this works for symbols loaded in shared libraries, but
>>>> // does NOT seem to find symbols declared in the executable.
>>>> if (auto Addr =
>>>> RTDyldMemoryManager::getSymbolAddressInProcess(name)) {
>>>> return RuntimeDyld::SymbolInfo(Addr, JITSymbolFlags::Exported);
>>>> }
>>>> },
>>>> [](const std::string &S) { return nullptr; }
>>>> );
>>>> }
>>>>
>>>> Here the getPointerToGlobalMapping function looks in a uint64 StringMap
>>>> into which values are added via the addGlobalMapping functions.
>>>>
>>>>
>>>> This approach seems to be working, but my question is do you suppose
>>>> there any are issues with such an approach?
>>>>
>>>> The troubling thing is why doesn’t  RTDyldMemoryManager::
>>>> getSymbolAddressInProcess(name)) return an address for a symbol that
>>>> is defined in either a static library, or in the executable itself.
>>>>
>>>> If this approach is correct, in adding the global values to the
>>>> context, and looking them up the lambda resolver, in addition to looking up
>>>> external symbols, and considering that the ORC kaleidoscope examples do in
>>>> fact allow external function calls (which are broken currently), should
>>>> they be fixed with this approach?
>>>>
>>>> _______________________________________________
>>>> LLVM Developers mailing list
>>>> llvm-dev at lists.llvm.org         http://llvm.cs.uiuc.edu
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>
>>>>
>>>
>>
>>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150820/0348a129/attachment.html>