1 //===--- AMDGPUMacroFusion.cpp - AMDGPU Macro Fusion ----------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file This file contains the AMDGPU implementation of the DAG scheduling
11 ///  mutation to pair instructions back to back.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPUMacroFusion.h"
16 #include "AMDGPUSubtarget.h"
17 #include "SIInstrInfo.h"
18 
19 #include "llvm/CodeGen/MacroFusion.h"
20 
21 using namespace llvm;
22 
23 namespace {
24 
25 /// \brief Check if the instr pair, FirstMI and SecondMI, should be fused
26 /// together. Given SecondMI, when FirstMI is unspecified, then check if
27 /// SecondMI may be part of a fused pair at all.
28 static bool shouldScheduleAdjacent(const TargetInstrInfo &TII_,
29                                    const TargetSubtargetInfo &TSI,
30                                    const MachineInstr *FirstMI,
31                                    const MachineInstr &SecondMI) {
32   const SIInstrInfo &TII = static_cast<const SIInstrInfo&>(TII_);
33 
34   switch (SecondMI.getOpcode()) {
35   case AMDGPU::V_ADDC_U32_e64:
36   case AMDGPU::V_SUBB_U32_e64:
37   case AMDGPU::V_CNDMASK_B32_e64: {
38     // Try to cluster defs of condition registers to their uses. This improves
39     // the chance VCC will be available which will allow shrinking to VOP2
40     // encodings.
41     if (!FirstMI)
42       return true;
43 
44     const MachineOperand *Src2 = TII.getNamedOperand(SecondMI,
45                                                      AMDGPU::OpName::src2);
46     return FirstMI->definesRegister(Src2->getReg());
47   }
48   default:
49     return false;
50   }
51 
52   return false;
53 }
54 
55 } // end namespace
56 
57 
58 namespace llvm {
59 
60 std::unique_ptr<ScheduleDAGMutation> createAMDGPUMacroFusionDAGMutation () {
61   return createMacroFusionDAGMutation(shouldScheduleAdjacent);
62 }
63 
64 } // end namespace llvm
65