[llvm-commits] [llvm] r62286 - in /llvm/trunk: include/llvm/CodeGen/DwarfWriter.h lib/CodeGen/AsmPrinter/DwarfWriter.cpp lib/CodeGen/SelectionDAG/FastISel.cpp lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp test/DebugInfo/2009-01-15-dbg_declare.ll

[Chris Lattner]

On Jan 16, 2009, at 9:55 AM, Evan Cheng wrote:

> Hi Devang,
>
> Is this the right fix or just a work around? Should llvm-gcc be  
> generating these invalid declare intrinsics in the first place? It  
> seems like this will increase compile time. It could be noticeable  
> when it's in fastisel mode, no?
>
> +    DIDescriptor DI(GV);
> +    // Check current version. Allow Version6 for now.
> +    unsigned Version = DI.getVersion();
> +    if (Version != DIDescriptor::Version7 && Version !=  
> DIDescriptor::Version6)
> +      return false;
>
> I know DIDescriptor is light weight. But is it necessary to create  
> the descriptor just to check the version number? Can you add a  
> helper function to DebugInfo that examines GV and returns its  
> corresponding version?

This should be cheap.  DIDescriptor just holds the pointer to the  
global.  This should inline into:

    unsigned getVersion() const {
       return getUnsignedField(0) & VersionMask / 
*VersionMask=0xffff0000*/;
     }

->

     unsigned getUnsignedField(unsigned Elt) const {
       return (unsigned)getUInt64Field(Elt);
     }

->

     uint64_t getUInt64Field(unsigned Elt) const;

->

uint64_t DIDescriptor::getUInt64Field(unsigned Elt) const {
   if (GV == 0) return 0;
   Constant *C = GV->getInitializer();
   if (C == 0 || Elt >= C->getNumOperands())
     return 0;
   if (ConstantInt *CI = dyn_cast<ConstantInt>(C->getOperand(Elt)))
     return CI->getZExtValue();
   return 0;
}

Which is pretty much what you have to do in any case.  DIDescriptor  
shouldn't add over head here as long as it isn't "new"d.

-Chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20090116/826f495e/attachment.html>