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>
39 extern "C" void LLVMInitializeR600Target() {
40 // Register the target
41 RegisterTargetMachine<AMDGPUTargetMachine> X(TheAMDGPUTarget);
44 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
45 return new ScheduleDAGMI(C, new R600SchedStrategy());
48 static MachineSchedRegistry
49 SchedCustomRegistry("r600", "Run R600's custom scheduler",
50 createR600MachineScheduler);
52 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, StringRef TT,
53 StringRef CPU, StringRef FS,
54 TargetOptions Options,
55 Reloc::Model RM, CodeModel::Model CM,
56 CodeGenOpt::Level OptLevel
59 LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
60 Subtarget(TT, CPU, FS),
61 Layout(Subtarget.getDataLayout()),
62 FrameLowering(TargetFrameLowering::StackGrowsUp, 16 // Stack Alignment
65 InstrItins(&Subtarget.getInstrItineraryData()) {
66 // TLInfo uses InstrInfo so it must be initialized after.
67 if (Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
68 InstrInfo.reset(new R600InstrInfo(*this));
69 TLInfo.reset(new R600TargetLowering(*this));
71 InstrInfo.reset(new SIInstrInfo(*this));
72 TLInfo.reset(new SITargetLowering(*this));
77 AMDGPUTargetMachine::~AMDGPUTargetMachine() {
81 class AMDGPUPassConfig : public TargetPassConfig {
83 AMDGPUPassConfig(AMDGPUTargetMachine *TM, PassManagerBase &PM)
84 : TargetPassConfig(TM, PM) {}
86 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
87 return getTM<AMDGPUTargetMachine>();
90 virtual ScheduleDAGInstrs *
91 createMachineScheduler(MachineSchedContext *C) const {
92 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
93 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
94 return createR600MachineScheduler(C);
98 virtual bool addPreISel();
99 virtual bool addInstSelector();
100 virtual bool addPreRegAlloc();
101 virtual bool addPostRegAlloc();
102 virtual bool addPreSched2();
103 virtual bool addPreEmitPass();
105 } // End of anonymous namespace
107 TargetPassConfig *AMDGPUTargetMachine::createPassConfig(PassManagerBase &PM) {
108 return new AMDGPUPassConfig(this, PM);
111 //===----------------------------------------------------------------------===//
112 // AMDGPU Analysis Pass Setup
113 //===----------------------------------------------------------------------===//
115 void AMDGPUTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
116 // Add first the target-independent BasicTTI pass, then our AMDGPU pass. This
117 // allows the AMDGPU pass to delegate to the target independent layer when
119 PM.add(createBasicTargetTransformInfoPass(this));
120 PM.add(createAMDGPUTargetTransformInfoPass(this));
124 AMDGPUPassConfig::addPreISel() {
125 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
126 addPass(createFlattenCFGPass());
127 if (ST.IsIRStructurizerEnabled() ||
128 ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS)
129 addPass(createStructurizeCFGPass());
130 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
131 addPass(createSITypeRewriter());
132 addPass(createSIAnnotateControlFlowPass());
134 addPass(createR600TextureIntrinsicsReplacer());
139 bool AMDGPUPassConfig::addInstSelector() {
140 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
142 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
143 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
144 // This callbacks this pass uses are not implemented yet on SI.
145 addPass(createAMDGPUIndirectAddressingPass(*TM));
150 bool AMDGPUPassConfig::addPreRegAlloc() {
151 addPass(createAMDGPUConvertToISAPass(*TM));
152 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
154 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
155 addPass(createR600VectorRegMerger(*TM));
157 addPass(createSIFixSGPRCopiesPass(*TM));
162 bool AMDGPUPassConfig::addPostRegAlloc() {
163 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
165 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
166 addPass(createSIInsertWaits(*TM));
171 bool AMDGPUPassConfig::addPreSched2() {
172 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
174 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
175 addPass(createR600EmitClauseMarkers(*TM));
176 addPass(&IfConverterID);
177 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
178 addPass(createR600ClauseMergePass(*TM));
182 bool AMDGPUPassConfig::addPreEmitPass() {
183 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
184 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
185 addPass(createAMDGPUCFGStructurizerPass(*TM));
186 addPass(createR600ExpandSpecialInstrsPass(*TM));
187 addPass(&FinalizeMachineBundlesID);
188 addPass(createR600Packetizer(*TM));
189 addPass(createR600ControlFlowFinalizer(*TM));
191 addPass(createSILowerControlFlowPass(*TM));