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 "AMDGPUTargetTransformInfo.h"
19 #include "R600ISelLowering.h"
20 #include "R600InstrInfo.h"
21 #include "R600MachineScheduler.h"
22 #include "SIISelLowering.h"
23 #include "SIInstrInfo.h"
24 #include "llvm/Analysis/Passes.h"
25 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
26 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
27 #include "llvm/CodeGen/MachineModuleInfo.h"
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/IR/Verifier.h"
30 #include "llvm/MC/MCAsmInfo.h"
31 #include "llvm/PassManager.h"
32 #include "llvm/Support/TargetRegistry.h"
33 #include "llvm/Support/raw_os_ostream.h"
34 #include "llvm/Transforms/IPO.h"
35 #include "llvm/Transforms/Scalar.h"
36 #include <llvm/CodeGen/Passes.h>
40 extern "C" void LLVMInitializeR600Target() {
41 // Register the target
42 RegisterTargetMachine<AMDGPUTargetMachine> X(TheAMDGPUTarget);
43 RegisterTargetMachine<GCNTargetMachine> Y(TheGCNTarget);
46 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
47 return new ScheduleDAGMILive(C, make_unique<R600SchedStrategy>());
50 static MachineSchedRegistry
51 SchedCustomRegistry("r600", "Run R600's custom scheduler",
52 createR600MachineScheduler);
54 static std::string computeDataLayout(StringRef TT) {
56 std::string Ret = "e-p:32:32";
58 if (Triple.getArch() == Triple::amdgcn) {
59 // 32-bit private, local, and region pointers. 64-bit global and constant.
60 Ret += "-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32-p24:64:64";
63 Ret += "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256"
64 "-v512:512-v1024:1024-v2048:2048-n32:64";
69 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, StringRef TT,
70 StringRef CPU, StringRef FS,
71 TargetOptions Options, Reloc::Model RM,
73 CodeGenOpt::Level OptLevel)
74 : LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
75 DL(computeDataLayout(TT)),
76 TLOF(new TargetLoweringObjectFileELF()),
77 Subtarget(TT, CPU, FS, *this), IntrinsicInfo() {
78 setRequiresStructuredCFG(true);
82 AMDGPUTargetMachine::~AMDGPUTargetMachine() {
87 class AMDGPUPassConfig : public TargetPassConfig {
89 AMDGPUPassConfig(AMDGPUTargetMachine *TM, PassManagerBase &PM)
90 : TargetPassConfig(TM, PM) {}
92 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
93 return getTM<AMDGPUTargetMachine>();
97 createMachineScheduler(MachineSchedContext *C) const override {
98 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
99 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
100 return createR600MachineScheduler(C);
104 void addIRPasses() override;
105 void addCodeGenPrepare() override;
106 bool addPreISel() override;
107 bool addInstSelector() override;
108 void addPreRegAlloc() override;
109 void addPostRegAlloc() override;
110 void addPreSched2() override;
111 void addPreEmitPass() override;
113 } // End of anonymous namespace
115 TargetPassConfig *AMDGPUTargetMachine::createPassConfig(PassManagerBase &PM) {
116 return new AMDGPUPassConfig(this, PM);
119 //===----------------------------------------------------------------------===//
121 //===----------------------------------------------------------------------===//
123 TargetIRAnalysis AMDGPUTargetMachine::getTargetIRAnalysis() {
124 return TargetIRAnalysis(
125 [this](Function &F) { return TargetTransformInfo(AMDGPUTTIImpl(this)); });
128 void AMDGPUPassConfig::addIRPasses() {
129 // Function calls are not supported, so make sure we inline everything.
130 addPass(createAMDGPUAlwaysInlinePass());
131 addPass(createAlwaysInlinerPass());
132 // We need to add the barrier noop pass, otherwise adding the function
133 // inlining pass will cause all of the PassConfigs passes to be run
134 // one function at a time, which means if we have a nodule with two
135 // functions, then we will generate code for the first function
136 // without ever running any passes on the second.
137 addPass(createBarrierNoopPass());
138 TargetPassConfig::addIRPasses();
141 void AMDGPUPassConfig::addCodeGenPrepare() {
142 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
143 if (ST.isPromoteAllocaEnabled()) {
144 addPass(createAMDGPUPromoteAlloca(ST));
145 addPass(createSROAPass());
148 TargetPassConfig::addCodeGenPrepare();
152 AMDGPUPassConfig::addPreISel() {
153 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
154 addPass(createFlattenCFGPass());
155 if (ST.IsIRStructurizerEnabled())
156 addPass(createStructurizeCFGPass());
157 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
158 addPass(createSinkingPass());
159 addPass(createSITypeRewriter());
160 addPass(createSIAnnotateControlFlowPass());
162 addPass(createR600TextureIntrinsicsReplacer());
167 bool AMDGPUPassConfig::addInstSelector() {
168 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
170 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
172 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
173 addPass(createSILowerI1CopiesPass());
174 addPass(createSIFixSGPRCopiesPass(*TM));
175 addPass(createSIFoldOperandsPass());
181 void AMDGPUPassConfig::addPreRegAlloc() {
182 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
184 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
185 addPass(createR600VectorRegMerger(*TM));
187 if (getOptLevel() > CodeGenOpt::None && ST.loadStoreOptEnabled()) {
188 // Don't do this with no optimizations since it throws away debug info by
189 // merging nonadjacent loads.
191 // This should be run after scheduling, but before register allocation. It
192 // also need extra copies to the address operand to be eliminated.
193 initializeSILoadStoreOptimizerPass(*PassRegistry::getPassRegistry());
194 insertPass(&MachineSchedulerID, &SILoadStoreOptimizerID);
197 addPass(createSIShrinkInstructionsPass(), false);
198 addPass(createSIFixSGPRLiveRangesPass(), false);
202 void AMDGPUPassConfig::addPostRegAlloc() {
203 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
205 if (ST.getGeneration() > AMDGPUSubtarget::NORTHERN_ISLANDS) {
206 addPass(createSIPrepareScratchRegs(), false);
207 addPass(createSIShrinkInstructionsPass(), false);
211 void AMDGPUPassConfig::addPreSched2() {
212 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
214 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
215 addPass(createR600EmitClauseMarkers(), false);
216 if (ST.isIfCvtEnabled())
217 addPass(&IfConverterID, false);
218 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
219 addPass(createR600ClauseMergePass(*TM), false);
220 if (ST.getGeneration() >= AMDGPUSubtarget::SOUTHERN_ISLANDS) {
221 addPass(createSIInsertWaits(*TM), false);
225 void AMDGPUPassConfig::addPreEmitPass() {
226 const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
227 if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
228 addPass(createAMDGPUCFGStructurizerPass(), false);
229 addPass(createR600ExpandSpecialInstrsPass(*TM), false);
230 addPass(&FinalizeMachineBundlesID, false);
231 addPass(createR600Packetizer(*TM), false);
232 addPass(createR600ControlFlowFinalizer(*TM), false);
234 addPass(createSILowerControlFlowPass(*TM), false);
239 //===----------------------------------------------------------------------===//
240 // GCN Target Machine (SI+)
241 //===----------------------------------------------------------------------===//
243 GCNTargetMachine::GCNTargetMachine(const Target &T, StringRef TT, StringRef FS,
244 StringRef CPU, TargetOptions Options, Reloc::Model RM,
245 CodeModel::Model CM, CodeGenOpt::Level OL) :
246 AMDGPUTargetMachine(T, TT, FS, CPU, Options, RM, CM, OL) { }