[PATCH] R600/SI: Teach moveToVALU how to handle some SMRD instructions

[Tom Stellard]

---
 lib/Target/R600/SIInstrInfo.cpp   | 48 ++++++++++++++++++++++++++++++++++++++-
 lib/Target/R600/SIInstrInfo.h     |  2 ++
 test/CodeGen/R600/salu-to-valu.ll | 28 +++++++++++++++++++++++
 3 files changed, 77 insertions(+), 1 deletion(-)

diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
index 96eeea5..d358c0d 100644
--- a/lib/Target/R600/SIInstrInfo.cpp
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -548,6 +548,9 @@ unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
   case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e32;
   case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e32;
   case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e32;
+  case AMDGPU::S_LOAD_DWORD_SGPR: return AMDGPU::BUFFER_LOAD_DWORD_ADDR64;
+  case AMDGPU::S_LOAD_DWORDX2_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX2_ADDR64;
+  case AMDGPU::S_LOAD_DWORDX4_SGPR: return AMDGPU::BUFFER_LOAD_DWORDX4_ADDR64;
   }
 }
 
@@ -910,6 +913,44 @@ void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
   }
 }
 
+void SIInstrInfo::moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const {
+  MachineBasicBlock *MBB = MI->getParent();
+  switch (MI->getOpcode()) {
+    case AMDGPU::S_LOAD_DWORD_SGPR:
+    case AMDGPU::S_LOAD_DWORDX2_SGPR:
+    case AMDGPU::S_LOAD_DWORDX4_SGPR:
+      unsigned NewOpcode = getVALUOp(*MI);
+      unsigned Offset = MI->getOperand(2).getReg();
+
+
+      unsigned SRsrc = MRI.createVirtualRegister(&AMDGPU::SReg_128RegClass);
+      unsigned DWord0 = Offset;
+      unsigned DWord1 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
+      unsigned DWord2 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
+      unsigned DWord3 = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
+
+      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord1)
+              .addImm(0);
+      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord2)
+              .addImm(AMDGPU::RSRC_DATA_FORMAT & 0xFFFFFFFF);
+      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::S_MOV_B32), DWord3)
+              .addImm(AMDGPU::RSRC_DATA_FORMAT >> 32);
+      BuildMI(*MBB, MI, MI->getDebugLoc(), get(AMDGPU::REG_SEQUENCE), SRsrc)
+              .addReg(DWord0)
+              .addImm(AMDGPU::sub0)
+              .addReg(DWord1)
+              .addImm(AMDGPU::sub1)
+              .addReg(DWord2)
+              .addImm(AMDGPU::sub2)
+              .addReg(DWord3)
+              .addImm(AMDGPU::sub3);
+     MI->setDesc(get(NewOpcode));
+     MI->getOperand(2).setReg(MI->getOperand(1).getReg());
+     MI->getOperand(1).setReg(SRsrc);
+     MI->addOperand(*MBB->getParent(), MachineOperand::CreateImm(0));
+  }
+}
+
 void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
   SmallVector<MachineInstr *, 128> Worklist;
   Worklist.push_back(&TopInst);
@@ -920,9 +961,15 @@ void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
     MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
 
     unsigned Opcode = Inst->getOpcode();
+    unsigned NewOpcode = getVALUOp(*Inst);
 
     // Handle some special cases
     switch (Opcode) {
+    default:
+      if (isSMRD(Inst->getOpcode())) {
+        moveSMRDToVALU(Inst, MRI);
+      }
+      break;
     case AMDGPU::S_MOV_B64: {
       DebugLoc DL = Inst->getDebugLoc();
 
@@ -973,7 +1020,6 @@ void SIInstrInfo::moveToVALU(MachineInstr &TopInst) const {
       llvm_unreachable("Moving this op to VALU not implemented");
     }
 
-    unsigned NewOpcode = getVALUOp(*Inst);
     if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
       // We cannot move this instruction to the VALU, so we should try to
       // legalize its operands instead.
diff --git a/lib/Target/R600/SIInstrInfo.h b/lib/Target/R600/SIInstrInfo.h
index 044cf37..18ff80b 100644
--- a/lib/Target/R600/SIInstrInfo.h
+++ b/lib/Target/R600/SIInstrInfo.h
@@ -137,6 +137,8 @@ public:
   /// create new instruction and insert them before \p MI.
   void legalizeOperands(MachineInstr *MI) const;
 
+  void moveSMRDToVALU(MachineInstr *MI, MachineRegisterInfo &MRI) const;
+
   /// \brief Replace this instruction's opcode with the equivalent VALU
   /// opcode.  This function will also move the users of \p MI to the
   /// VALU if necessary.
diff --git a/test/CodeGen/R600/salu-to-valu.ll b/test/CodeGen/R600/salu-to-valu.ll
index e461bf9..3fd0db3 100644
--- a/test/CodeGen/R600/salu-to-valu.ll
+++ b/test/CodeGen/R600/salu-to-valu.ll
@@ -46,3 +46,31 @@ declare i32 @llvm.r600.read.tidig.x() #1
 declare i32 @llvm.r600.read.tidig.y() #1
 
 attributes #1 = { nounwind readnone }
+
+; Test moving an SMRD instruction to the VALU
+
+; CHECK-LABEL: @smrd_valu
+; CHECK: BUFFER_LOAD_DWORD [[OUT:v[0-9]+]]
+; CHECK: BUFFER_STORE_DWORD [[OUT]]
+
+define void @smrd_valu(i32 addrspace(2)* addrspace(1)* %in, i32 %a, i32 addrspace(1)* %out) {
+entry:
+  %0 = icmp ne i32 %a, 0
+  br i1 %0, label %if, label %else
+
+if:
+  %1 = load i32 addrspace(2)* addrspace(1)* %in
+  br label %endif
+
+else:
+  %2 = getelementptr i32 addrspace(2)* addrspace(1)* %in
+  %3 = load i32 addrspace(2)* addrspace(1)* %2
+  br label %endif
+
+endif:
+  %4 = phi i32 addrspace(2)*  [%1, %if], [%3, %else]
+  %5 = getelementptr i32 addrspace(2)* %4, i32 3000
+  %6 = load i32 addrspace(2)* %5
+  store i32 %6, i32 addrspace(1)* %out
+  ret void
+}
-- 
1.8.1.5