llvm dev - Jan 2009 - [LLVMdev] Newbie question: Getting info about JIT-compiled function

Andrew Haley wrote:
> Jeff Kuskin wrote:
> > Apologies if this is a FAQ.
> >
> > I am using the LLVM JIT facility on an x86_64 platform.  I generate
> > IR for a single function using IRBuilder(), use the
> > FunctionPassManager to do some optimization passes, and then call
> > ExecutionEngine::getPointerToFunction() to get a native-code version
> > of the function.  Everything works fine so far.
> >
> > Two questions:
> >
> > Question (1): How can I get the size of the native function?  I see
> > that the information is stored in the __jitSymbolTable structure in
> > JITEmitter.cpp, but I don't see any interface to query this symbol
> > table.  I see AddFunctionToSymbolTable() and
> > RemoveFunctionFromSymbolTable(), but not any equivalent of
> > GetFunctionFromSymbolTable().  Is there a supported way to get at
> > the JitSymbolEntry for a function?  If not, is there an alternate
> > way to determine the size (in bytes) of the native function?
> 
> The answer seems to be "create your own memory manager and define
> endFunctionBody()".  If you have a look at class MyJITMemoryManager in
>
http://icedtea.classpath.org/hg/icedtea6/file/0507a324ec22/ports/hotspot/src/share/vm/shark/sharkBuilder.hpp
you'll see how I did it.


I took a look and it makes sense (though it does seem a bit heavyweight).
I'll give it a try.

> > Question (2): Does LLVM have any routines for printing the
> > disassembly of a native code function?  I see that lli can do this,
> > but I would much prefer an LLVM library routine that I can call from
> > my own code rather than having to write the bitcode to a file and
> > then invoke lli.  Suggestions?
> 
> This keeps changing, but on x86 with 4.2 you do this to enable
> assembly dumps:
> 
> std::vector<const char*> Args;
> Args.push_back(""); // program name
> Args.push_back("-debug-only=" "x86-emitter");
> cl::ParseCommandLineOptions(Args.size()-1, (char**)&Args[0]);
I must be missing something here.  I added this code and it is executed
before any function is compiled.  But I still don't see the x86 asm
printed.  I build the function, run the FunctionPassManager,
and then call getPointerToFunction().  All works fine.  But I don't
see any printing when the FunctionPassManager runs, and even if I
call f->print() (f being the Function*) after the FunctionPassManager
runs, I still see only the LLVM asm but not x86 asm.

What am I doing wrong?

Thanks!

-- Jeff

On Jan 28, 2009, at 6:37 AM, Jeff Kuskin wrote:
> Andrew Haley wrote:
>> Jeff Kuskin wrote:
>>> Apologies if this is a FAQ.
>>>
>>> I am using the LLVM JIT facility on an x86_64 platform.  I generate
>>> IR for a single function using IRBuilder(), use the
>>> FunctionPassManager to do some optimization passes, and then call
>>> ExecutionEngine::getPointerToFunction() to get a native-code
version
>>> of the function.  Everything works fine so far.
>>>
>>> Two questions:
>>>
>>> Question (1): How can I get the size of the native function?  I see
>>> that the information is stored in the __jitSymbolTable structure in
>>> JITEmitter.cpp, but I don't see any interface to query this
symbol
>>> table.  I see AddFunctionToSymbolTable() and
>>> RemoveFunctionFromSymbolTable(), but not any equivalent of
>>> GetFunctionFromSymbolTable().  Is there a supported way to get at
>>> the JitSymbolEntry for a function?  If not, is there an alternate
>>> way to determine the size (in bytes) of the native function?
>>
>> The answer seems to be "create your own memory manager and define
>> endFunctionBody()".  If you have a look at class
MyJITMemoryManager
>> in
>>
http://icedtea.classpath.org/hg/icedtea6/file/0507a324ec22/ports/hotspot/src/share/vm/shark/sharkBuilder.hpp
> you'll see how I did it.
>
>
> I took a look and it makes sense (though it does seem a bit  
> heavyweight).
> I'll give it a try.
>
>
>>> Question (2): Does LLVM have any routines for printing the
>>> disassembly of a native code function?  I see that lli can do this,
>>> but I would much prefer an LLVM library routine that I can call
from
>>> my own code rather than having to write the bitcode to a file and
>>> then invoke lli.  Suggestions?
>>
>> This keeps changing, but on x86 with 4.2 you do this to enable
>> assembly dumps:
>>
>> std::vector<const char*> Args;
>> Args.push_back(""); // program name
>> Args.push_back("-debug-only=" "x86-emitter");
>> cl::ParseCommandLineOptions(Args.size()-1, (char**)&Args[0]);
>
> I must be missing something here.  I added this code and it is  
> executed
> before any function is compiled.  But I still don't see the x86 asm
> printed.  I build the function, run the FunctionPassManager,
> and then call getPointerToFunction().  All works fine.  But I don't
> see any printing when the FunctionPassManager runs, and even if I
> call f->print() (f being the Function*) after the FunctionPassManager
> runs, I still see only the LLVM asm but not x86 asm.
Instead of "x86-emitter", try "jit".

Evan
>
>
> What am I doing wrong?
>
> Thanks!
>
> -- Jeff
>
>
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev