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/CodeGen/MachineFunctionAnalysis.h"
25 #include "llvm/CodeGen/MachineModuleInfo.h"
26 #include "llvm/CodeGen/Passes.h"
27 #include "llvm/IR/Verifier.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 ScheduleDAGMILive(C, make_unique<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, Reloc::Model RM,
55 CodeGenOpt::Level OptLevel)
56 : LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
57 Subtarget(TT, CPU, FS, *this), IntrinsicInfo() {
58 setRequiresStructuredCFG(true);
62 AMDGPUTargetMachine::~AMDGPUTargetMachine() {
66 class AMDGPUPassConfig : public TargetPassConfig {
68 AMDGPUPassConfig(AMDGPUTargetMachine *TM, PassManagerBase &PM)
69 : TargetPassConfig(TM, PM) {}
71 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
72 return getTM<AMDGPUTargetMachine>();
76 createMachineScheduler(MachineSchedContext *C) const override {
77 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
78 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
79 return createR600MachineScheduler(C);
83 void addCodeGenPrepare() override;
84 bool addPreISel() override;
85 bool addInstSelector() override;
86 bool addPreRegAlloc() override;
87 bool addPostRegAlloc() override;
88 bool addPreSched2() override;
89 bool addPreEmitPass() override;
91 } // End of anonymous namespace
93 TargetPassConfig *AMDGPUTargetMachine::createPassConfig(PassManagerBase &PM) {
94 return new AMDGPUPassConfig(this, PM);
97 //===----------------------------------------------------------------------===//
98 // AMDGPU Analysis Pass Setup
99 //===----------------------------------------------------------------------===//
101 void AMDGPUTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
102 // Add first the target-independent BasicTTI pass, then our AMDGPU pass. This
103 // allows the AMDGPU pass to delegate to the target independent layer when
105 PM.add(createBasicTargetTransformInfoPass(this));
106 PM.add(createAMDGPUTargetTransformInfoPass(this));
109 void AMDGPUPassConfig::addCodeGenPrepare() {
110 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
111 if (ST.isPromoteAllocaEnabled()) {
112 addPass(createAMDGPUPromoteAlloca(ST));
113 addPass(createSROAPass());
116 TargetPassConfig::addCodeGenPrepare();
120 AMDGPUPassConfig::addPreISel() {
121 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
122 addPass(createFlattenCFGPass());
123 if (ST.IsIRStructurizerEnabled())
124 addPass(createStructurizeCFGPass());
125 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
126 addPass(createSinkingPass());
127 addPass(createSITypeRewriter());
128 addPass(createSIAnnotateControlFlowPass());
130 addPass(createR600TextureIntrinsicsReplacer());
135 bool AMDGPUPassConfig::addInstSelector() {
136 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
137 addPass(createSILowerI1CopiesPass());
141 bool AMDGPUPassConfig::addPreRegAlloc() {
142 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
144 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
145 addPass(createR600VectorRegMerger(*TM));
147 addPass(createSIFixSGPRCopiesPass(*TM));
148 // SIFixSGPRCopies can generate a lot of duplicate instructions,
149 // so we need to run MachineCSE afterwards.
150 addPass(&MachineCSEID);
152 if (getOptLevel() > CodeGenOpt::None && ST.loadStoreOptEnabled()) {
153 // Don't do this with no optimizations since it throws away debug info by
154 // merging nonadjacent loads.
156 // This should be run after scheduling, but before register allocation. It
157 // also need extra copies to the address operand to be eliminated.
158 initializeSILoadStoreOptimizerPass(*PassRegistry::getPassRegistry());
159 insertPass(&MachineSchedulerID, &SILoadStoreOptimizerID);
162 addPass(createSIShrinkInstructionsPass());
163 addPass(createSIFixSGPRLiveRangesPass());
168 bool AMDGPUPassConfig::addPostRegAlloc() {
169 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
171 addPass(createSIShrinkInstructionsPass());
172 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
173 addPass(createSIInsertWaits(*TM));
178 bool AMDGPUPassConfig::addPreSched2() {
179 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
181 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
182 addPass(createR600EmitClauseMarkers());
183 if (ST.isIfCvtEnabled())
184 addPass(&IfConverterID);
185 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
186 addPass(createR600ClauseMergePass(*TM));
190 bool AMDGPUPassConfig::addPreEmitPass() {
191 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
192 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
193 addPass(createAMDGPUCFGStructurizerPass());
194 addPass(createR600ExpandSpecialInstrsPass(*TM));
195 addPass(&FinalizeMachineBundlesID);
196 addPass(createR600Packetizer(*TM));
197 addPass(createR600ControlFlowFinalizer(*TM));
199 addPass(createSILowerControlFlowPass(*TM));