2 files changed, 4 insertions, 34 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index e8de28263f9..3851ee8968b 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -2189,22 +2189,7 @@ void AMDGPUDAGToDAGISel::SelectDS_GWS(SDNode *N, unsigned IntrID) {
     glueCopyToM0(N, SDValue(M0Base, 0));
   }
 
-  SDValue V0;
   SDValue Chain = N->getOperand(0);
-  SDValue Glue;
-  if (HasVSrc) {
-    SDValue VSrc0 = N->getOperand(2);
-
-    // The manual doesn't mention this, but it seems only v0 works.
-    V0 = CurDAG->getRegister(AMDGPU::VGPR0, MVT::i32);
-
-    SDValue CopyToV0 = CurDAG->getCopyToReg(
-      N->getOperand(0), SL, V0, VSrc0,
-      N->getOperand(N->getNumOperands() - 1));
-    Chain = CopyToV0;
-    Glue = CopyToV0.getValue(1);
-  }
-
   SDValue OffsetField = CurDAG->getTargetConstant(ImmOffset, SL, MVT::i32);
 
   // TODO: Can this just be removed from the instruction?
@@ -2213,14 +2198,11 @@ void AMDGPUDAGToDAGISel::SelectDS_GWS(SDNode *N, unsigned IntrID) {
   const unsigned Opc = gwsIntrinToOpcode(IntrID);
   SmallVector<SDValue, 5> Ops;
   if (HasVSrc)
-    Ops.push_back(V0);
+    Ops.push_back(N->getOperand(2));
   Ops.push_back(OffsetField);
   Ops.push_back(GDS);
   Ops.push_back(Chain);
 
-  if (HasVSrc)
-    Ops.push_back(Glue);
-
   SDNode *Selected = CurDAG->SelectNodeTo(N, Opc, N->getVTList(), Ops);
   CurDAG->setNodeMemRefs(cast<MachineSDNode>(Selected), {MMO});
 }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index bee25c9d184..a14a8929aa4 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -3096,12 +3096,13 @@ SITargetLowering::emitGWSMemViolTestLoop(MachineInstr &MI,
   MachineBasicBlock *RemainderBB;
   const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
 
-  MachineBasicBlock::iterator Prev = std::prev(MI.getIterator());
+  // Apparently kill flags are only valid if the def is in the same block?
+  if (MachineOperand *Src = TII->getNamedOperand(MI, AMDGPU::OpName::data0))
+    Src->setIsKill(false);
 
   std::tie(LoopBB, RemainderBB) = splitBlockForLoop(MI, *BB, true);
 
   MachineBasicBlock::iterator I = LoopBB->end();
-  MachineOperand *Src = TII->getNamedOperand(MI, AMDGPU::OpName::data0);
 
   const unsigned EncodedReg = AMDGPU::Hwreg::encodeHwreg(
     AMDGPU::Hwreg::ID_TRAPSTS, AMDGPU::Hwreg::OFFSET_MEM_VIOL, 1);
@@ -3111,19 +3112,6 @@ SITargetLowering::emitGWSMemViolTestLoop(MachineInstr &MI,
     .addImm(0)
     .addImm(EncodedReg);
 
-  // This is a pain, but we're not allowed to have physical register live-ins
-  // yet. Insert a pair of copies if the VGPR0 hack is necessary.
-  if (Src && TargetRegisterInfo::isPhysicalRegister(Src->getReg())) {
-    unsigned Data0 = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
-    BuildMI(*BB, std::next(Prev), DL, TII->get(AMDGPU::COPY), Data0)
-      .add(*Src);
-
-    BuildMI(*LoopBB, LoopBB->begin(), DL, TII->get(AMDGPU::COPY), Src->getReg())
-      .addReg(Data0);
-
-    MRI.setSimpleHint(Data0, Src->getReg());
-  }
-
   bundleInstWithWaitcnt(MI);
 
   unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);