[Libclc-dev] [PATCH v3 1/1] R600: Add new intrinsic to read work dimensions

Tom Stellard tom at stellard.net
Thu Aug 7 09:11:02 PDT 2014


On Thu, Aug 07, 2014 at 12:01:00PM -0400, Jan Vesely wrote:
> On Wed, 2014-08-06 at 15:21 -0700, Matt Arsenault wrote:
> > On 08/06/2014 03:08 PM, Jan Vesely wrote:
> > > +  case Intrinsic::r600_read_workdim: {
> > > +    const size_t arg_size = DAG.getMachineFunction().getFunction()->arg_size();
> > arg_size() returns the number of arguments, not their actual size. You 
> > can't assume every argument is 4 bytes. There could be larger types, 
> > vectors, or structs
> 
> ah, right. I guess I'll have to repeat most of the argument magic from
> clover/llvm/invocation.cpp
> 

The way I would do this is to add a field to AMDGPUMachineFunctionInfo
called something like ABIArgOffset, and then set in
SITargetLowering::LowerFormalArguments() where we compute the offset for
each argument something like:

   for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {

     const ISD::InputArg &Arg = Ins[i];
     if (Skipped & (1 << i)) {
       InVals.push_back(DAG.getUNDEF(Arg.VT));
       continue;
     }

     CCValAssign &VA = ArgLocs[ArgIdx++];
     EVT VT = VA.getLocVT();

     if (VA.isMemLoc()) {
       VT = Ins[i].VT;
       EVT MemVT = Splits[i].VT;
+      unsigned Offset = 36 + VA.getLocMemOffset();
       // The first 36 bytes of the input buffer contains information
       // about
       // thread group and global sizes.
       SDValue Arg = LowerParameter(DAG, VT, MemVT,  DL, DAG.getRoot(),
+                                   Offset,
-                                   36 + VA.getLocMemOffset(),
                                    Ins[i].Flags.isSExt());
       InVals.push_back(Arg);
+      MFI->ABIArgOffset = Offset + MemVT.getSizeInBits() / 8
       continue;
     }


Then can you can use this computed offset when you lower get_dim.

-Tom

> > > +    return LowerParameter(DAG, VT, VT, DL, DAG.getEntryNode(), 36 + (arg_size * 4), false);
> > > +  }
> > 
> 
> -- 
> Jan Vesely <jan.vesely at rutgers.edu>






More information about the Libclc-dev mailing list