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 static std::string computeDataLayout(const AMDGPUSubtarget &ST) {
53 std::string Ret = "e";
58 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS)
61 Ret += "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256"
62 "-v512:512-v1024:1024-v2048:2048-n32:64";
67 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, StringRef TT,
68 StringRef CPU, StringRef FS,
69 TargetOptions Options,
70 Reloc::Model RM, CodeModel::Model CM,
71 CodeGenOpt::Level OptLevel
74 LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
75 Subtarget(TT, CPU, FS),
76 Layout(computeDataLayout(Subtarget)),
77 FrameLowering(TargetFrameLowering::StackGrowsUp,
78 64 * 16 // Maximum stack alignment (long16)
81 InstrItins(&Subtarget.getInstrItineraryData()) {
82 // TLInfo uses InstrInfo so it must be initialized after.
83 if (Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
84 InstrInfo.reset(new R600InstrInfo(*this));
85 TLInfo.reset(new R600TargetLowering(*this));
87 InstrInfo.reset(new SIInstrInfo(*this));
88 TLInfo.reset(new SITargetLowering(*this));
90 setRequiresStructuredCFG(true);
94 AMDGPUTargetMachine::~AMDGPUTargetMachine() {
98 class AMDGPUPassConfig : public TargetPassConfig {
100 AMDGPUPassConfig(AMDGPUTargetMachine *TM, PassManagerBase &PM)
101 : TargetPassConfig(TM, PM) {}
103 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
104 return getTM<AMDGPUTargetMachine>();
107 virtual ScheduleDAGInstrs *
108 createMachineScheduler(MachineSchedContext *C) const {
109 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
110 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
111 return createR600MachineScheduler(C);
115 virtual bool addPreISel();
116 virtual bool addInstSelector();
117 virtual bool addPreRegAlloc();
118 virtual bool addPostRegAlloc();
119 virtual bool addPreSched2();
120 virtual bool addPreEmitPass();
122 } // End of anonymous namespace
124 TargetPassConfig *AMDGPUTargetMachine::createPassConfig(PassManagerBase &PM) {
125 return new AMDGPUPassConfig(this, PM);
128 //===----------------------------------------------------------------------===//
129 // AMDGPU Analysis Pass Setup
130 //===----------------------------------------------------------------------===//
132 void AMDGPUTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
133 // Add first the target-independent BasicTTI pass, then our AMDGPU pass. This
134 // allows the AMDGPU pass to delegate to the target independent layer when
136 PM.add(createBasicTargetTransformInfoPass(this));
137 PM.add(createAMDGPUTargetTransformInfoPass(this));
141 AMDGPUPassConfig::addPreISel() {
142 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
143 addPass(createFlattenCFGPass());
144 if (ST.IsIRStructurizerEnabled())
145 addPass(createStructurizeCFGPass());
146 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
147 addPass(createSinkingPass());
148 addPass(createSITypeRewriter());
149 addPass(createSIAnnotateControlFlowPass());
151 addPass(createR600TextureIntrinsicsReplacer());
156 bool AMDGPUPassConfig::addInstSelector() {
157 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
161 bool AMDGPUPassConfig::addPreRegAlloc() {
162 addPass(createAMDGPUConvertToISAPass(*TM));
163 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
165 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
166 addPass(createR600VectorRegMerger(*TM));
168 addPass(createSIFixSGPRCopiesPass(*TM));
173 bool AMDGPUPassConfig::addPostRegAlloc() {
174 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
176 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
177 addPass(createSIInsertWaits(*TM));
182 bool AMDGPUPassConfig::addPreSched2() {
183 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
185 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
186 addPass(createR600EmitClauseMarkers());
187 if (ST.isIfCvtEnabled())
188 addPass(&IfConverterID);
189 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
190 addPass(createR600ClauseMergePass(*TM));
194 bool AMDGPUPassConfig::addPreEmitPass() {
195 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
196 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
197 addPass(createAMDGPUCFGStructurizerPass());
198 addPass(createR600ExpandSpecialInstrsPass(*TM));
199 addPass(&FinalizeMachineBundlesID);
200 addPass(createR600Packetizer(*TM));
201 addPass(createR600ControlFlowFinalizer(*TM));
203 addPass(createSILowerControlFlowPass(*TM));