1 //===-- AMDGPU.h - MachineFunction passes hw codegen --------------*- C++ -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
9 //===----------------------------------------------------------------------===//
11 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPU_H
12 #define LLVM_LIB_TARGET_AMDGPU_AMDGPU_H
14 #include "llvm/Target/TargetMachine.h"
18 class AMDGPUTargetMachine;
20 class GCNTargetMachine;
29 FunctionPass *createR600VectorRegMerger(TargetMachine &tm);
30 FunctionPass *createR600ExpandSpecialInstrsPass(TargetMachine &tm);
31 FunctionPass *createR600EmitClauseMarkers();
32 FunctionPass *createR600ClauseMergePass(TargetMachine &tm);
33 FunctionPass *createR600Packetizer(TargetMachine &tm);
34 FunctionPass *createR600ControlFlowFinalizer(TargetMachine &tm);
35 FunctionPass *createAMDGPUCFGStructurizerPass();
38 FunctionPass *createSITypeRewriter();
39 FunctionPass *createSIAnnotateControlFlowPass();
40 FunctionPass *createSIFoldOperandsPass();
41 FunctionPass *createSIPeepholeSDWAPass();
42 FunctionPass *createSILowerI1CopiesPass();
43 FunctionPass *createSIShrinkInstructionsPass();
44 FunctionPass *createSILoadStoreOptimizerPass(TargetMachine &tm);
45 FunctionPass *createSIWholeQuadModePass();
46 FunctionPass *createSIFixControlFlowLiveIntervalsPass();
47 FunctionPass *createSIFixSGPRCopiesPass();
48 FunctionPass *createSIDebuggerInsertNopsPass();
49 FunctionPass *createSIInsertWaitsPass();
50 FunctionPass *createAMDGPUCodeGenPreparePass(const GCNTargetMachine *TM = nullptr);
52 ModulePass *createAMDGPUAnnotateKernelFeaturesPass(const TargetMachine *TM = nullptr);
53 void initializeAMDGPUAnnotateKernelFeaturesPass(PassRegistry &);
54 extern char &AMDGPUAnnotateKernelFeaturesID;
56 ModulePass *createAMDGPULowerIntrinsicsPass();
57 void initializeAMDGPULowerIntrinsicsPass(PassRegistry &);
58 extern char &AMDGPULowerIntrinsicsID;
60 void initializeSIFoldOperandsPass(PassRegistry &);
61 extern char &SIFoldOperandsID;
63 void initializeSIPeepholeSDWAPass(PassRegistry &);
64 extern char &SIPeepholeSDWAID;
66 void initializeSIShrinkInstructionsPass(PassRegistry&);
67 extern char &SIShrinkInstructionsID;
69 void initializeSIFixSGPRCopiesPass(PassRegistry &);
70 extern char &SIFixSGPRCopiesID;
72 void initializeSIFixVGPRCopiesPass(PassRegistry &);
73 extern char &SIFixVGPRCopiesID;
75 void initializeSILowerI1CopiesPass(PassRegistry &);
76 extern char &SILowerI1CopiesID;
78 void initializeSILoadStoreOptimizerPass(PassRegistry &);
79 extern char &SILoadStoreOptimizerID;
81 void initializeSIWholeQuadModePass(PassRegistry &);
82 extern char &SIWholeQuadModeID;
84 void initializeSILowerControlFlowPass(PassRegistry &);
85 extern char &SILowerControlFlowID;
87 void initializeSIInsertSkipsPass(PassRegistry &);
88 extern char &SIInsertSkipsPassID;
90 void initializeSIOptimizeExecMaskingPass(PassRegistry &);
91 extern char &SIOptimizeExecMaskingID;
93 // Passes common to R600 and SI
94 FunctionPass *createAMDGPUPromoteAlloca(const TargetMachine *TM = nullptr);
95 void initializeAMDGPUPromoteAllocaPass(PassRegistry&);
96 extern char &AMDGPUPromoteAllocaID;
98 Pass *createAMDGPUStructurizeCFGPass();
99 FunctionPass *createAMDGPUISelDag(TargetMachine &TM,
100 CodeGenOpt::Level OptLevel);
101 ModulePass *createAMDGPUAlwaysInlinePass();
102 ModulePass *createAMDGPUOpenCLImageTypeLoweringPass();
103 FunctionPass *createAMDGPUAnnotateUniformValues();
105 ModulePass* createAMDGPUUnifyMetadataPass();
106 void initializeAMDGPUUnifyMetadataPass(PassRegistry&);
107 extern char &AMDGPUUnifyMetadataID;
109 void initializeSIFixControlFlowLiveIntervalsPass(PassRegistry&);
110 extern char &SIFixControlFlowLiveIntervalsID;
112 void initializeAMDGPUAnnotateUniformValuesPass(PassRegistry&);
113 extern char &AMDGPUAnnotateUniformValuesPassID;
115 void initializeAMDGPUCodeGenPreparePass(PassRegistry&);
116 extern char &AMDGPUCodeGenPrepareID;
118 void initializeSIAnnotateControlFlowPass(PassRegistry&);
119 extern char &SIAnnotateControlFlowPassID;
121 void initializeSIDebuggerInsertNopsPass(PassRegistry&);
122 extern char &SIDebuggerInsertNopsID;
124 void initializeSIInsertWaitsPass(PassRegistry&);
125 extern char &SIInsertWaitsID;
127 void initializeAMDGPUUnifyDivergentExitNodesPass(PassRegistry&);
128 extern char &AMDGPUUnifyDivergentExitNodesID;
130 ImmutablePass *createAMDGPUAAWrapperPass();
131 void initializeAMDGPUAAWrapperPassPass(PassRegistry&);
133 Target &getTheAMDGPUTarget();
134 Target &getTheGCNTarget();
139 TI_SCRATCH_RSRC_DWORD0,
140 TI_SCRATCH_RSRC_DWORD1,
141 TI_SCRATCH_RSRC_DWORD2,
142 TI_SCRATCH_RSRC_DWORD3
146 } // End namespace llvm
148 /// OpenCL uses address spaces to differentiate between
149 /// various memory regions on the hardware. On the CPU
150 /// all of the address spaces point to the same memory,
151 /// however on the GPU, each address space points to
152 /// a separate piece of memory that is unique from other
153 /// memory locations.
155 // The following address space values depend on the triple environment.
156 unsigned PRIVATE_ADDRESS; ///< Address space for private memory.
157 unsigned CONSTANT_ADDRESS; ///< Address space for constant memory (VTX2)
158 unsigned FLAT_ADDRESS; ///< Address space for flat memory.
159 unsigned REGION_ADDRESS; ///< Address space for region memory.
161 // The maximum value for flat, generic, local, private, constant and region.
162 const static unsigned MAX_COMMON_ADDRESS = 5;
164 const static unsigned GLOBAL_ADDRESS = 1; ///< Address space for global memory (RAT0, VTX0).
165 const static unsigned LOCAL_ADDRESS = 3; ///< Address space for local memory.
166 const static unsigned PARAM_D_ADDRESS = 6; ///< Address space for direct addressible parameter memory (CONST0)
167 const static unsigned PARAM_I_ADDRESS = 7; ///< Address space for indirect addressible parameter memory (VTX1)
169 // Do not re-order the CONSTANT_BUFFER_* enums. Several places depend on this
170 // order to be able to dynamically index a constant buffer, for example:
172 // ConstantBufferAS = CONSTANT_BUFFER_0 + CBIdx
174 const static unsigned CONSTANT_BUFFER_0 = 8;
175 const static unsigned CONSTANT_BUFFER_1 = 9;
176 const static unsigned CONSTANT_BUFFER_2 = 10;
177 const static unsigned CONSTANT_BUFFER_3 = 11;
178 const static unsigned CONSTANT_BUFFER_4 = 12;
179 const static unsigned CONSTANT_BUFFER_5 = 13;
180 const static unsigned CONSTANT_BUFFER_6 = 14;
181 const static unsigned CONSTANT_BUFFER_7 = 15;
182 const static unsigned CONSTANT_BUFFER_8 = 16;
183 const static unsigned CONSTANT_BUFFER_9 = 17;
184 const static unsigned CONSTANT_BUFFER_10 = 18;
185 const static unsigned CONSTANT_BUFFER_11 = 19;
186 const static unsigned CONSTANT_BUFFER_12 = 20;
187 const static unsigned CONSTANT_BUFFER_13 = 21;
188 const static unsigned CONSTANT_BUFFER_14 = 22;
189 const static unsigned CONSTANT_BUFFER_15 = 23;
191 // Some places use this if the address space can't be determined.
192 const static unsigned UNKNOWN_ADDRESS_SPACE = ~0u;
197 AMDGPUAS getAMDGPUAS(const Module &M);
198 AMDGPUAS getAMDGPUAS(const TargetMachine &TM);
199 AMDGPUAS getAMDGPUAS(Triple T);
200 } // namespace AMDGPU