1 //===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief The AMDGPU target machine contains all of the hardware specific
12 /// information needed to emit code for R600 and SI GPUs.
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUTargetMachine.h"
18 #include "R600ISelLowering.h"
19 #include "R600InstrInfo.h"
20 #include "R600MachineScheduler.h"
21 #include "SIISelLowering.h"
22 #include "SIInstrInfo.h"
23 #include "llvm/Analysis/Passes.h"
24 #include "llvm/Analysis/Verifier.h"
25 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
26 #include "llvm/CodeGen/MachineModuleInfo.h"
27 #include "llvm/CodeGen/Passes.h"
28 #include "llvm/MC/MCAsmInfo.h"
29 #include "llvm/PassManager.h"
30 #include "llvm/Support/TargetRegistry.h"
31 #include "llvm/Support/raw_os_ostream.h"
32 #include "llvm/Transforms/IPO.h"
33 #include "llvm/Transforms/Scalar.h"
34 #include <llvm/CodeGen/Passes.h>
38 extern "C" void LLVMInitializeR600Target() {
39 // Register the target
40 RegisterTargetMachine<AMDGPUTargetMachine> X(TheAMDGPUTarget);
43 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
44 return new ScheduleDAGMI(C, new R600SchedStrategy());
47 static MachineSchedRegistry
48 SchedCustomRegistry("r600", "Run R600's custom scheduler",
49 createR600MachineScheduler);
51 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, StringRef TT,
52 StringRef CPU, StringRef FS,
53 TargetOptions Options,
54 Reloc::Model RM, CodeModel::Model CM,
55 CodeGenOpt::Level OptLevel
58 LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
59 Subtarget(TT, CPU, FS),
60 Layout(Subtarget.getDataLayout()),
61 FrameLowering(TargetFrameLowering::StackGrowsUp, 16 // Stack Alignment
64 InstrItins(&Subtarget.getInstrItineraryData()) {
65 // TLInfo uses InstrInfo so it must be initialized after.
66 if (Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
67 InstrInfo.reset(new R600InstrInfo(*this));
68 TLInfo.reset(new R600TargetLowering(*this));
70 InstrInfo.reset(new SIInstrInfo(*this));
71 TLInfo.reset(new SITargetLowering(*this));
76 AMDGPUTargetMachine::~AMDGPUTargetMachine() {
80 class AMDGPUPassConfig : public TargetPassConfig {
82 AMDGPUPassConfig(AMDGPUTargetMachine *TM, PassManagerBase &PM)
83 : TargetPassConfig(TM, PM) {}
85 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
86 return getTM<AMDGPUTargetMachine>();
89 virtual ScheduleDAGInstrs *
90 createMachineScheduler(MachineSchedContext *C) const {
91 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
92 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
93 return createR600MachineScheduler(C);
97 virtual bool addPreISel();
98 virtual bool addInstSelector();
99 virtual bool addPreRegAlloc();
100 virtual bool addPostRegAlloc();
101 virtual bool addPreSched2();
102 virtual bool addPreEmitPass();
104 } // End of anonymous namespace
106 TargetPassConfig *AMDGPUTargetMachine::createPassConfig(PassManagerBase &PM) {
107 return new AMDGPUPassConfig(this, PM);
110 //===----------------------------------------------------------------------===//
111 // AMDGPU Analysis Pass Setup
112 //===----------------------------------------------------------------------===//
114 void AMDGPUTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
115 // Add first the target-independent BasicTTI pass, then our AMDGPU pass. This
116 // allows the AMDGPU pass to delegate to the target independent layer when
118 PM.add(createBasicTargetTransformInfoPass(this));
119 PM.add(createAMDGPUTargetTransformInfoPass(this));
123 AMDGPUPassConfig::addPreISel() {
124 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
125 addPass(createFlattenCFGPass());
126 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
127 addPass(createSITypeRewriter());
128 addPass(createStructurizeCFGPass());
129 addPass(createSIAnnotateControlFlowPass());
131 addPass(createR600TextureIntrinsicsReplacer());
136 bool AMDGPUPassConfig::addInstSelector() {
137 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
139 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
140 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
141 // This callbacks this pass uses are not implemented yet on SI.
142 addPass(createAMDGPUIndirectAddressingPass(*TM));
147 bool AMDGPUPassConfig::addPreRegAlloc() {
148 addPass(createAMDGPUConvertToISAPass(*TM));
149 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
151 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
152 addPass(createR600VectorRegMerger(*TM));
154 addPass(createSIFixSGPRCopiesPass(*TM));
159 bool AMDGPUPassConfig::addPostRegAlloc() {
160 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
162 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
163 addPass(createSIInsertWaits(*TM));
168 bool AMDGPUPassConfig::addPreSched2() {
169 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
171 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
172 addPass(createR600EmitClauseMarkers(*TM));
173 addPass(&IfConverterID);
174 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
175 addPass(createR600ClauseMergePass(*TM));
179 bool AMDGPUPassConfig::addPreEmitPass() {
180 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
181 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
182 addPass(createAMDGPUCFGStructurizerPass(*TM));
183 addPass(createR600ExpandSpecialInstrsPass(*TM));
184 addPass(&FinalizeMachineBundlesID);
185 addPass(createR600Packetizer(*TM));
186 addPass(createR600ControlFlowFinalizer(*TM));
188 addPass(createSILowerControlFlowPass(*TM));