[llvm] r241960 - AMDGPU: Fix chains for memory ops dependent on argument loads

[Matt Arsenault]

Author: arsenm
Date: Fri Jul 10 17:51:36 2015
New Revision: 241960

URL: http://llvm.org/viewvc/llvm-project?rev=241960&view=rev
Log:
AMDGPU: Fix chains for memory ops dependent on argument loads

Most loads and stores are derived from pointers derived from
a kernel argument load inserted during argument lowering.
This was just using the EntryToken chain for the argument loads,
and any users of these loads were also on the EntryToken chain.

Return the chain of the lowered argument load so that dependent loads
end up on the correct chain.

No test since I'm not aware of any case where this actually
broke.

Modified:
    llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp

Modified: llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp?rev=241960&r1=241959&r2=241960&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp Fri Jul 10 17:51:36 2015
@@ -452,7 +452,12 @@ SDValue SITargetLowering::LowerParameter
                                true, // isNonTemporal
                                true, // isInvariant
                                Align); // Alignment
-    return DAG.getNode(ISD::FP16_TO_FP, SL, VT, Load);
+    SDValue Ops[] = {
+      DAG.getNode(ISD::FP16_TO_FP, SL, VT, Load),
+      Load.getValue(1)
+    };
+
+    return DAG.getMergeValues(Ops, SL);
   }
 
   ISD::LoadExtType ExtTy = Signed ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
@@ -570,6 +575,8 @@ SDValue SITargetLowering::LowerFormalArg
 
   AnalyzeFormalArguments(CCInfo, Splits);
 
+  SmallVector<SDValue, 16> Chains;
+
   for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
 
     const ISD::InputArg &Arg = Ins[i];
@@ -590,6 +597,7 @@ SDValue SITargetLowering::LowerFormalArg
       // thread group and global sizes.
       SDValue Arg = LowerParameter(DAG, VT, MemVT,  DL, Chain,
                                    Offset, Ins[i].Flags.isSExt());
+      Chains.push_back(Arg.getValue(1));
 
       const PointerType *ParamTy =
         dyn_cast<PointerType>(FType->getParamType(Ins[i].getOrigArgIndex()));
@@ -615,7 +623,8 @@ SDValue SITargetLowering::LowerFormalArg
       Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
                                      &AMDGPU::SReg_64RegClass);
       Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
-      InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
+      SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
+      InVals.push_back(Copy);
       continue;
     }
 
@@ -635,7 +644,9 @@ SDValue SITargetLowering::LowerFormalArg
       for (unsigned j = 1; j != NumElements; ++j) {
         Reg = ArgLocs[ArgIdx++].getLocReg();
         Reg = MF.addLiveIn(Reg, RC);
-        Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
+
+        SDValue Copy = DAG.getCopyFromReg(Chain, DL, Reg, VT);
+        Regs.push_back(Copy);
       }
 
       // Fill up the missing vector elements
@@ -654,7 +665,11 @@ SDValue SITargetLowering::LowerFormalArg
         AMDGPU::SGPR_32RegClass.begin(), AMDGPU::SGPR_32RegClass.getNumRegs()));
     Info->ScratchOffsetReg = AMDGPU::SGPR_32RegClass.getRegister(ScratchIdx);
   }
-  return Chain;
+
+  if (Chains.empty())
+    return Chain;
+
+  return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
 }
 
 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(