R600/SI: Remove explicit m0 operand from s_sendmsg

Instead add m0 as an implicit operand.  This allows us to avoid using
the M0Reg register class and eliminates a number of unnecessary spills
when using s_sendmsg instructions.  This impacts one shader in the
shader-db:

SGPRS: 48 -> 40 (-16.67 %)
VGPRS: 112 -> 108 (-3.57 %)
Code Size: 40132 -> 38796 (-3.33 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Scratch: 2048 -> 0 (-100.00 %) bytes per wave

llvm-svn: 237133

6 files changed, 36 insertions, 8 deletions

diff --git a/llvm/lib/Target/R600/AMDGPUISelLowering.cpp b/llvm/lib/Target/R600/AMDGPUISelLowering.cpp
index 9c987c28d7f..ceb7e151959 100644
--- a/llvm/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -2677,6 +2677,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
   NODE_NAME_CASE(CONST_DATA_PTR)
   case AMDGPUISD::FIRST_MEM_OPCODE_NUMBER: break;
+  NODE_NAME_CASE(SENDMSG)
   NODE_NAME_CASE(STORE_MSKOR)
   NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
   case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
diff --git a/llvm/lib/Target/R600/AMDGPUISelLowering.h b/llvm/lib/Target/R600/AMDGPUISelLowering.h
index fe657d5d731..a6bfbbd1572 100644
--- a/llvm/lib/Target/R600/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/R600/AMDGPUISelLowering.h
@@ -288,6 +288,7 @@ enum NodeType : unsigned {
   BUILD_VERTICAL_VECTOR,
   /// Pointer to the start of the shader's constant data.
   CONST_DATA_PTR,
+  SENDMSG,
   FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
   STORE_MSKOR,
   LOAD_CONSTANT,
diff --git a/llvm/lib/Target/R600/AMDGPUInstrInfo.td b/llvm/lib/Target/R600/AMDGPUInstrInfo.td
index d72cb1d7f8c..e251cb48e0d 100644
--- a/llvm/lib/Target/R600/AMDGPUInstrInfo.td
+++ b/llvm/lib/Target/R600/AMDGPUInstrInfo.td
@@ -219,6 +219,10 @@ def AMDGPUmad_i24 : SDNode<"AMDGPUISD::MAD_I24", AMDGPUDTIntTernaryOp,
   []
 >;
 
+def AMDGPUsendmsg : SDNode<"AMDGPUISD::SENDMSG",
+                    SDTypeProfile<0, 1, [SDTCisInt<0>]>,
+                    [SDNPHasChain, SDNPInGlue]>;
+
 //===----------------------------------------------------------------------===//
 // Flow Control Profile Types
 //===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp
index 6ff2a9689e0..56214cd07dc 100644
--- a/llvm/lib/Target/R600/SIISelLowering.cpp
+++ b/llvm/lib/Target/R600/SIISelLowering.cpp
@@ -838,6 +838,23 @@ SDValue SITargetLowering::LowerGlobalAddress(AMDGPUMachineFunction *MFI,
   return DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Lo, Hi);
 }
 
+SDValue SITargetLowering::copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL,
+                                   SDValue V) const {
+  // 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.
+}
+
 SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                   SelectionDAG &DAG) const {
   MachineFunction &MF = DAG.getMachineFunction();
@@ -940,12 +957,18 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
 SDValue SITargetLowering::LowerINTRINSIC_VOID(SDValue Op,
                                               SelectionDAG &DAG) const {
   MachineFunction &MF = DAG.getMachineFunction();
+  SDLoc DL(Op);
   SDValue Chain = Op.getOperand(0);
   unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
 
   switch (IntrinsicID) {
+  case AMDGPUIntrinsic::SI_sendmsg: {
+    Chain = copyToM0(DAG, Chain, DL, Op.getOperand(3));
+    SDValue Glue = Chain.getValue(1);
+    return DAG.getNode(AMDGPUISD::SENDMSG, DL, MVT::Other, Chain,
+                       Op.getOperand(2), Glue);
+  }
   case AMDGPUIntrinsic::SI_tbuffer_store: {
-    SDLoc DL(Op);
     SDValue Ops[] = {
       Chain,
       Op.getOperand(2),
diff --git a/llvm/lib/Target/R600/SIISelLowering.h b/llvm/lib/Target/R600/SIISelLowering.h
index a6bc7c607fa..a95354c3881 100644
--- a/llvm/lib/Target/R600/SIISelLowering.h
+++ b/llvm/lib/Target/R600/SIISelLowering.h
@@ -117,6 +117,7 @@ public:
   std::pair<unsigned, const TargetRegisterClass *> getRegForInlineAsmConstraint(
                                    const TargetRegisterInfo *TRI,
                                    const std::string &Constraint, MVT VT) const override;
+  SDValue copyToM0(SelectionDAG &DAG, SDValue Chain, SDLoc DL, SDValue V) const;
 };
 
 } // End namespace llvm
diff --git a/llvm/lib/Target/R600/SIInstructions.td b/llvm/lib/Target/R600/SIInstructions.td
index 838b645f2ba..0449ba8cd18 100644
--- a/llvm/lib/Target/R600/SIInstructions.td
+++ b/llvm/lib/Target/R600/SIInstructions.td
@@ -488,13 +488,11 @@ def S_SETHALT : SOPP <0x0000000d, (ins i16imm:$simm16), "s_sethalt $simm16">;
 def S_SLEEP : SOPP <0x0000000e, (ins i16imm:$simm16), "s_sleep $simm16">;
 def S_SETPRIO : SOPP <0x0000000f, (ins i16imm:$sim16), "s_setprio $sim16">;
 
-let Uses = [EXEC] in {
-  def S_SENDMSG : SOPP <0x00000010, (ins SendMsgImm:$simm16, M0Reg:$m0), "s_sendmsg $simm16",
-      [(int_SI_sendmsg imm:$simm16, M0Reg:$m0)]
-  > {
-    let DisableEncoding = "$m0";
-  }
-} // End Uses = [EXEC]
+let Uses = [EXEC, M0] in {
+  def S_SENDMSG : SOPP <0x00000010, (ins SendMsgImm:$simm16), "s_sendmsg $simm16",
+      [(AMDGPUsendmsg (i32 imm:$simm16))]
+  >;
+} // End Uses = [EXEC, M0]
 
 def S_SENDMSGHALT : SOPP <0x00000011, (ins i16imm:$simm16), "s_sendmsghalt $simm16">;
 def S_TRAP : SOPP <0x00000012, (ins i16imm:$simm16), "s_trap $simm16">;