llvm dev - Oct 2011 - [LLVMdev] Expected behavior of eliminateFrameIndex() on dbg_value machine instructions

I'm investigating a bug associated with debug information that manifests 
itself in the XCore backend (PR11105). I'd like to understand what the 
expected behavior of eliminateFrameIndex() is when it is called on a 
dbg_value machine instruction.

Currently the XCore target replaces the frame index with the frame 
register and sets the next operand to the byte offset from the frame 
register. A quick glance at some of the other targets suggests this is 
the right thing to do (for example 
ARMBaseRegisterInfo::eliminateFrameIndex and 
Thumb1RegisterInfo::eliminateFrameIndex appear to handle dbg_value in 
exactly the same way).

Unfortunately this results in an assertion firing in 
MachineFrameInfo::getObjectOffset(). This is due to 
CompileUnit::constructVariableDIE() passing the second operand of the 
dbg_value instruction (set to the byte offset) to 
getFrameIndexReference() which expects a frame index as an argument.

Am I doing something wrong in eliminateFrameIndex()? Is this a bug in 
CompileUnit::constructVariableDIE()? Is there something else I'm missing?

Thanks,

Richard

-- 
Richard Osborne | XMOS
http://www.xmos.com

On Oct 10, 2011, at 10:26 AM, Richard Osborne wrote:
> I'm investigating a bug associated with debug information that
manifests
> itself in the XCore backend (PR11105). I'd like to understand what the 
> expected behavior of eliminateFrameIndex() is when it is called on a 
> dbg_value machine instruction.
That is up to the target.

The TII::emitFrameIndexDebugValue() hook is called to insert DBG_VALUE
instructions referring to values on the stack.

The target's eliminateFrameIndex() should be able to handle any DBG_VALUE
instructions created by this hook.

If you don't implement the hook, you will get DBG_VALUE instructions with
just a FrameIndex operand followed by the mandatory Offset and Metadata
operands.

/jakob

On 10/10/11 19:19, Jakob Stoklund Olesen wrote:
> On Oct 10, 2011, at 10:26 AM, Richard Osborne wrote:
>> I'm investigating a bug associated with debug information that
manifests
>> itself in the XCore backend (PR11105). I'd like to understand what
the
>> expected behavior of eliminateFrameIndex() is when it is called on a
>> dbg_value machine instruction.
> That is up to the target.
>
> The TII::emitFrameIndexDebugValue() hook is called to insert DBG_VALUE
instructions referring to values on the stack.
Thanks for pointing this out. Looking at the code in 
constructVariableDIE it seems to use the number of operands to determine 
if the DBG_VALUE is target specific or not. It is considered target 
specific if it does not have exactly 3 operands in which case the 
getDebugValueLocation hook is called. Implementing both 
TII::emitFrameIndexDebugValue() and AsmPrinter::getDebugValueLocation() 
in the XCore backend seems to fix the problem I was seeing.

I'm still confused about what the following code in 
CompileUnit::constructVariableDIE is trying to do:

   if (const MachineInstr *DVInsn = DV->getMInsn()) {
     bool updated = false;
     if (DVInsn->getNumOperands() == 3) {
       if (DVInsn->getOperand(0).isReg()) {
         const MachineOperand RegOp = DVInsn->getOperand(0);
         const TargetRegisterInfo *TRI = Asm->TM.getRegisterInfo();
         if (DVInsn->getOperand(1).isImm() &&
             TRI->getFrameRegister(*Asm->MF) == RegOp.getReg()) {
           unsigned FrameReg = 0;
           const TargetFrameLowering *TFI = Asm->TM.getFrameLowering();
           int Offset              TFI->getFrameIndexReference(*Asm->MF,
                                         DVInsn->getOperand(1).getImm(),
                                         FrameReg);
           MachineLocation Location(FrameReg, Offset);
           addVariableAddress(DV, VariableDie, Location);

If the intention is to handle the case where the target doesn't 
implement emitFrameIndexDebugValue(), then there is a mismatch between 
the operands it expects and the operands added in 
UserValue::insertDebugValue(). I've attached a patch which fixes this. 
Does this look correct to you?
> The target's eliminateFrameIndex() should be able to handle any
DBG_VALUE instructions created by this hook.
>
> If you don't implement the hook, you will get DBG_VALUE instructions
with just a FrameIndex operand followed by the mandatory Offset and Metadata
operands.
This behavior doesn't match the comment above
TII::emitFrameIndexDebugValue() which says "For targets that do not support
this the debug info is simply lost". In UserValue::insertDebugValue in
LiveDebugVariables.cpp if emitFrameIndexDebugValue returns 0 then it adds a
DBG_VALUE with the standard operands. However in other places where
emitFrameIndexDebugValue is called it appears to removes the debug value (e.g.
in InlineSpiller::spillAroundUses()). Should the comment be updated or is the
code in LiveDebugVariables.cpp doing the wrong thing?
> /jakob
-- 
Richard Osborne | XMOS
http://www.xmos.com

-------------- next part --------------
A non-text attachment was scrubbed...
Name: dbg_value.patch
Type: text/x-patch
Size: 1166 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20111011/b4dd9e1b/attachment.bin>