Line data    Source code

       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             : #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
      19             : 
      20             : #include "llvm/CodeGen/MacroFusion.h"
      21             : 
      22             : using namespace llvm;
      23             : 
      24             : namespace {
      25             : 
      26             : /// Check if the instr pair, FirstMI and SecondMI, should be fused
      27             : /// together. Given SecondMI, when FirstMI is unspecified, then check if
      28             : /// SecondMI may be part of a fused pair at all.
      29      379352 : static bool shouldScheduleAdjacent(const TargetInstrInfo &TII_,
      30             :                                    const TargetSubtargetInfo &TSI,
      31             :                                    const MachineInstr *FirstMI,
      32             :                                    const MachineInstr &SecondMI) {
      33             :   const SIInstrInfo &TII = static_cast<const SIInstrInfo&>(TII_);
      34             : 
      35      758704 :   switch (SecondMI.getOpcode()) {
      36       28234 :   case AMDGPU::V_ADDC_U32_e64:
      37             :   case AMDGPU::V_SUBB_U32_e64:
      38             :   case AMDGPU::V_CNDMASK_B32_e64: {
      39             :     // Try to cluster defs of condition registers to their uses. This improves
      40             :     // the chance VCC will be available which will allow shrinking to VOP2
      41             :     // encodings.
      42       28234 :     if (!FirstMI)
      43             :       return true;
      44             : 
      45             :     const MachineOperand *Src2 = TII.getNamedOperand(SecondMI,
      46             :                                                      AMDGPU::OpName::src2);
      47       18427 :     return FirstMI->definesRegister(Src2->getReg());
      48             :   }
      49             :   default:
      50             :     return false;
      51             :   }
      52             : 
      53             :   return false;
      54             : }
      55             : 
      56             : } // end namespace
      57             : 
      58             : 
      59             : namespace llvm {
      60             : 
      61       19517 : std::unique_ptr<ScheduleDAGMutation> createAMDGPUMacroFusionDAGMutation () {
      62       19517 :   return createMacroFusionDAGMutation(shouldScheduleAdjacent);
      63             : }
      64             : 
      65             : } // end namespace llvm