llvm dev - Oct 2010 - [LLVMdev] PowerPC : Assertion `MovePCtoLROffset && "MovePCtoLR not seen yet?"' failed.

Hi all,

I'm compiling current SVN HEAD on Linux/x86. The tests are failing
on PowerPC due to the following assertion failure :

    JITTests: PPCCodeEmitter.cpp:152: unsigned
int<unnamed>::PPCCodeEmitter::
    getMachineOpValue(const llvm::MachineInstr&, const
llvm::MachineOperand&)
    const: Assertion `MovePCtoLROffset && "MovePCtoLR not seen
yet?"' failed.

Is anyone working on the PowerPC backend?

Erik
-- 
----------------------------------------------------------------------
Erik de Castro Lopo
http://www.mega-nerd.com/

Erik de Castro Lopo <mle+cl <at> mega-nerd.com> writes:
> 
> Hi all,
> 
> I'm compiling current SVN HEAD on Linux/x86. The tests are failing
> on PowerPC due to the following assertion failure :
> 
>     JITTests: PPCCodeEmitter.cpp:152: unsigned
int<unnamed>::PPCCodeEmitter::
>     getMachineOpValue(const llvm::MachineInstr&, const
llvm::MachineOperand&)
>     const: Assertion `MovePCtoLROffset && "MovePCtoLR not seen
yet?"' failed.
> 
> Is anyone working on the PowerPC backend?
> 
> Erik
I'm seeing these same errors on powerpc-yellowdog-linux-gnu.
Causes failures on some LLVM :: ExecutionEngine and LLVM-Unit ::
ExecutionEngine/JIT/Release+Asserts/JITTests tests.

I'm not working on it, but I might be able to help find the problem.  What
this means is that you're generating code in PIC mode, and an object that
requires a PIC register to reference is being addressed, and no PIC register was
allocated.   The allocation was supposed to happen in PPCDAGtoDAGISel::Select
when the reference was processed, and a MovePCtoLR instruction inserted at that
time.  Somebody can probably figure it out by stepping through that.

On Oct 27, 2010, at 10:16 AMPDT, D. Johnson wrote:
> Erik de Castro Lopo <mle+cl <at> mega-nerd.com> writes:
> 
>> 
>> Hi all,
>> 
>> I'm compiling current SVN HEAD on Linux/x86. The tests are failing
>> on PowerPC due to the following assertion failure :
>> 
>>    JITTests: PPCCodeEmitter.cpp:152: unsigned
int<unnamed>::PPCCodeEmitter::
>>    getMachineOpValue(const llvm::MachineInstr&, const
llvm::MachineOperand&)
>>    const: Assertion `MovePCtoLROffset && "MovePCtoLR not
seen yet?"' failed.
>> 
>> Is anyone working on the PowerPC backend?
>> 
>> Erik
> 
> I'm seeing these same errors on powerpc-yellowdog-linux-gnu.
> Causes failures on some LLVM :: ExecutionEngine and LLVM-Unit ::
> ExecutionEngine/JIT/Release+Asserts/JITTests tests.
> 
> 
> 
> 
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Erik de Castro Lopo <mle+cl <at> mega-nerd.com> writes:
> 
> Hi all,
> 
> I'm compiling current SVN HEAD on Linux/x86. The tests are failing
> on PowerPC due to the following assertion failure :
> 
>     JITTests: PPCCodeEmitter.cpp:152: unsigned
int<unnamed>::PPCCodeEmitter::
>     getMachineOpValue(const llvm::MachineInstr&, const
llvm::MachineOperand&)
>     const: Assertion `MovePCtoLROffset && "MovePCtoLR not seen
yet?"' failed.
> 
> Is anyone working on the PowerPC backend?
> 
> Erik

I added build, host, target = ppc-unknown-linux-gnu and enabled libffi. One of
hese gets rid of the assert errors. I'm guessing the B,H,T addition.
Configure
script guessed ppc64.

D. Johnson wrote:
> I added build, host, target = ppc-unknown-linux-gnu and enabled libffi. One
of
> hese gets rid of the assert errors. I'm guessing the B,H,T addition.
Confirmed. Configuring with:

    ./configure --host=ppc-unknown-linux-gnu --build=ppc-unknown-linux-gnu \
              --target=ppc-unknown-linux-gnu

passed all tests.
> Configure script guessed ppc64.
I suspect that this is because you and I are both running 32 bit user space
on a 64 bit kernel. I'm running Debian Testing.


Erik
-- 
----------------------------------------------------------------------
Erik de Castro Lopo
http://www.mega-nerd.com/