Index: lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/SIISelLowering.cpp
+++ lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1080,10 +1080,18 @@
                            + StringRef(RegName) + "\"."));
 }
 
-MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
-    MachineInstr * MI, MachineBasicBlock * BB) const {
-
+MachineBasicBlock *SITargetLowering::EmitInstrWithCustomInserter(
+  MachineInstr *MI, MachineBasicBlock *BB) const {
   switch (MI->getOpcode()) {
+  case AMDGPU::SI_INIT_M0: {
+    const SIInstrInfo *TII =
+      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
+    BuildMI(*BB, MI->getIterator(), MI->getDebugLoc(),
+            TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
+      .addOperand(MI->getOperand(0));
+    MI->eraseFromParent();
+    break;
+  }
   case AMDGPU::BRANCH:
     return BB;
   case AMDGPU::GET_GROUPSTATICSIZE: {
@@ -1372,19 +1380,18 @@
 
 SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL,
                                    SDValue V) const {
+  // We can't use S_MOV_B32 directly, because there is no way to specify m0 as
+  // the destination register.
+  //
   // We can't use CopyToReg, because MachineCSE won't combine COPY instructions,
   // so we will end up with redundant moves to m0.
   //
-  // We can't use S_MOV_B32, because there is no way to specify m0 as the
-  // destination register.
-  //
-  // We have to use them both.  Machine cse will combine all the S_MOV_B32
-  // instructions and the register coalescer eliminate the extra copies.
-  SDNode *M0 = DAG.getMachineNode(AMDGPU::S_MOV_B32, DL, V.getValueType(), V);
-  return DAG.getCopyToReg(Chain, DL, DAG.getRegister(AMDGPU::M0, MVT::i32),
-                          SDValue(M0, 0), SDValue()); // Glue
-                                                      // A Null SDValue creates
-                                                      // a glue result.
+  // We use a pseudo to ensure we emit s_mov_b32 with m0 as the direct result.
+
+  // A Null SDValue creates a glue result.
+  SDNode *M0 = DAG.getMachineNode(AMDGPU::SI_INIT_M0, DL, MVT::Other, MVT::Glue,
+                                  V, Chain);
+  return SDValue(M0, 0);
 }
 
 SDValue SITargetLowering::lowerImplicitZextParam(SelectionDAG &DAG,
Index: lib/Target/AMDGPU/SIInstructions.td
===================================================================
--- lib/Target/AMDGPU/SIInstructions.td
+++ lib/Target/AMDGPU/SIInstructions.td
@@ -2014,7 +2014,23 @@
 
 } // End mayLoad = 1, mayStore = 1, hasSideEffects = 1
 
-let Uses = [EXEC], Defs = [EXEC,VCC,M0] in {
+// Used as an isel pseudo to directly emit initialization with an
+// s_mov_b32 rather than a copy of another initialized
+// register. MachineCSE skips copies, and we don't want to have to
+// fold operands before it runs.
+def SI_INIT_M0 : InstSI <
+  (outs),
+  (ins SSrc_32:$src), "", []> {
+  let Defs = [M0];
+  let usesCustomInserter = 1;
+  let isPseudo = 1;
+  let isCodeGenOnly = 1;
+  let isAsCheapAsAMove = 1;
+  let SALU = 1;
+  let isReMaterializable = 1;
+}
+
+let Uses = [EXEC], Defs = [EXEC, VCC, M0] in {
 
 class SI_INDIRECT_SRC<RegisterClass rc> : InstSI <
   (outs VGPR_32:$dst, SReg_64:$temp),