Change 6420645 causes SIGBUS during deoptimisation at a safepoint on 64bit-SPARC

[Vladimir Kozlov]

Thank you, Volker

I filed C2 bug since the fix is for .ad file and is not related to COOP:

6880034: SIGBUS during deoptimisation at a safepoint on 64bit-SPARC

Vladimir

Volker Simonis wrote:
> Hi Coleen,
> 
> did you filed/found a bug report for this issue. I can volunteer to
> open a bug for this and once there's a bug ID I could also submit a
> patch to fix it if that would help. I just would like to keep this
> issue tracked.
> 
> Regards,
> Volker
> 
> On 9/1/09, Tom Rodriguez <Thomas.Rodriguez at sun.com> wrote:
>>> The problem can be easily solved by switching back to the old
>>> implementation of frame::oopmapreg_to_location(), but I
>> assume there
>>> was a rational behind the change and that the new implementation is
>>> probably necessary for compressed oops (at least that's what the whole
>>> change was all about). So I dug a little further and found that in my
>>> opinion the root cause of the whole problem is the strange numbering
>>> of the 16 upper double registers in sparc.ad. They are defined as
>>> follows:
>>>
>>> reg_def R_D32x(SOC, SOC, Op_RegD,255, F32->as_VMReg());
>>> reg_def R_D32 (SOC, SOC, Op_RegD,  1, F32->as_VMReg()->next());
>>> reg_def R_D34x(SOC, SOC, Op_RegD,255, F34->as_VMReg());
>>> reg_def R_D34 (SOC, SOC, Op_RegD,  3, F34->as_VMReg()->next());
>>> ...
>>> reg_def R_D62x(SOC, SOC, Op_RegD,255, F62->as_VMReg());
>>> reg_def R_D62 (SOC, SOC, Op_RegD, 31, F62->as_VMReg()->next());
>>>
>>  I don't really know why it's done this way.  It's certainly not consistent
>> with the others.  If it all works better I'd be ok with correcting.
>>
>>
>>> PS: while I was hunting the error, I also stumbled across the
>>> following code in RegisterSaver::save_live_registers():
>>>
>>>  // Save all the FP registers
>>>  int offset = d00_offset;
>>>  for( int i=0; i<64; i+=2 ) {
>>>   FloatRegister f = as_FloatRegister(i);
>>>   __ stf(FloatRegisterImpl::D,  f, SP, offset+STACK_BIAS);
>>>
>> map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
>> f->as_VMReg());
>>>   if (true) {
>>>     map->set_callee_saved(VMRegImpl::stack2reg((offset +
>>> sizeof(float))>>2), f->as_VMReg()->next());
>>>   }
>>>   offset += sizeof(double);
>>>  }
>>>
>>  That would probably be ok too.
>>
>>  tom
>>
>>
>>> In my opinion, this could be changed to:
>>>
>>>  // Save all the FP registers
>>>  int offset = d00_offset;
>>>  for( int i=0; i<64; i+=2 ) {
>>>   FloatRegister f = as_FloatRegister(i);
>>>   __ stf(FloatRegisterImpl::D,  f, SP, offset+STACK_BIAS);
>>>
>> map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
>> f->as_VMReg());
>>>   if (i < 32) { // VS 2009-08-31: the 16 upper double registers
>>> can't be used as floats anyway
>>>     map->set_callee_saved(VMRegImpl::stack2reg((offset +
>>> sizeof(float))>>2), f->as_VMReg()->next());
>>>   }
>>>   offset += sizeof(double);
>>>  }
>>>
>>> because the 16 upper double registers can't be used as floats anyway.
>>> Again, I didn't found any regression in my few tests. What do you
>>> think?
>>> <DeoptTest.java>
>>>
>>