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 static std::string computeDataLayout(const AMDGPUSubtarget &ST) {
52 std::string Ret = "e-p:32:32";
55 // 32-bit local, and region pointers. 64-bit private, global, and constant.
56 Ret += "-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64";
59 Ret += "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256"
60 "-v512:512-v1024:1024-v2048:2048-n32:64";
65 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, StringRef TT,
66 StringRef CPU, StringRef FS,
67 TargetOptions Options,
68 Reloc::Model RM, CodeModel::Model CM,
69 CodeGenOpt::Level OptLevel
72 LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
73 Subtarget(TT, CPU, FS),
74 Layout(computeDataLayout(Subtarget)),
75 FrameLowering(TargetFrameLowering::StackGrowsUp,
76 64 * 16 // Maximum stack alignment (long16)
79 InstrItins(&Subtarget.getInstrItineraryData()) {
80 // TLInfo uses InstrInfo so it must be initialized after.
81 if (Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
82 TLInfo.reset(new R600TargetLowering(*this));
84 TLInfo.reset(new SITargetLowering(*this));
86 setRequiresStructuredCFG(true);
90 AMDGPUTargetMachine::~AMDGPUTargetMachine() {
94 class AMDGPUPassConfig : public TargetPassConfig {
96 AMDGPUPassConfig(AMDGPUTargetMachine *TM, PassManagerBase &PM)
97 : TargetPassConfig(TM, PM) {}
99 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
100 return getTM<AMDGPUTargetMachine>();
104 createMachineScheduler(MachineSchedContext *C) const override {
105 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
106 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
107 return createR600MachineScheduler(C);
111 virtual void addCodeGenPrepare();
112 bool addPreISel() override;
113 bool addInstSelector() override;
114 bool addPreRegAlloc() override;
115 bool addPostRegAlloc() override;
116 bool addPreSched2() override;
117 bool addPreEmitPass() override;
119 } // End of anonymous namespace
121 TargetPassConfig *AMDGPUTargetMachine::createPassConfig(PassManagerBase &PM) {
122 return new AMDGPUPassConfig(this, PM);
125 //===----------------------------------------------------------------------===//
126 // AMDGPU Analysis Pass Setup
127 //===----------------------------------------------------------------------===//
129 void AMDGPUTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
130 // Add first the target-independent BasicTTI pass, then our AMDGPU pass. This
131 // allows the AMDGPU pass to delegate to the target independent layer when
133 PM.add(createBasicTargetTransformInfoPass(this));
134 PM.add(createAMDGPUTargetTransformInfoPass(this));
137 void AMDGPUPassConfig::addCodeGenPrepare() {
138 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
139 if (ST.isPromoteAllocaEnabled()) {
140 addPass(createAMDGPUPromoteAlloca(ST));
141 addPass(createSROAPass());
144 TargetPassConfig::addCodeGenPrepare();
148 AMDGPUPassConfig::addPreISel() {
149 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
150 addPass(createFlattenCFGPass());
151 if (ST.IsIRStructurizerEnabled())
152 addPass(createStructurizeCFGPass());
153 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
154 addPass(createSinkingPass());
155 addPass(createSITypeRewriter());
156 addPass(createSIAnnotateControlFlowPass());
158 addPass(createR600TextureIntrinsicsReplacer());
163 bool AMDGPUPassConfig::addInstSelector() {
164 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
165 addPass(createSILowerI1CopiesPass());
169 bool AMDGPUPassConfig::addPreRegAlloc() {
170 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
172 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
173 addPass(createR600VectorRegMerger(*TM));
175 addPass(createSIFixSGPRCopiesPass(*TM));
176 // SIFixSGPRCopies can generate a lot of duplicate instructions,
177 // so we need to run MachineCSE afterwards.
178 addPass(&MachineCSEID);
179 addPass(createSIShrinkInstructionsPass());
180 initializeSIFixSGPRLiveRangesPass(*PassRegistry::getPassRegistry());
181 insertPass(&RegisterCoalescerID, &SIFixSGPRLiveRangesID);
186 bool AMDGPUPassConfig::addPostRegAlloc() {
187 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
189 addPass(createSIShrinkInstructionsPass());
190 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
191 addPass(createSIInsertWaits(*TM));
196 bool AMDGPUPassConfig::addPreSched2() {
197 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
199 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
200 addPass(createR600EmitClauseMarkers());
201 if (ST.isIfCvtEnabled())
202 addPass(&IfConverterID);
203 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
204 addPass(createR600ClauseMergePass(*TM));
208 bool AMDGPUPassConfig::addPreEmitPass() {
209 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
210 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
211 addPass(createAMDGPUCFGStructurizerPass());
212 addPass(createR600ExpandSpecialInstrsPass(*TM));
213 addPass(&FinalizeMachineBundlesID);
214 addPass(createR600Packetizer(*TM));
215 addPass(createR600ControlFlowFinalizer(*TM));
217 addPass(createSILowerControlFlowPass(*TM));