1 //===-- AArch64TargetMachine.cpp - Define TargetMachine for AArch64 -------===//
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 //===----------------------------------------------------------------------===//
14 #include "AArch64TargetMachine.h"
15 #include "llvm/CodeGen/Passes.h"
16 #include "llvm/CodeGen/RegAllocRegistry.h"
17 #include "llvm/IR/Function.h"
18 #include "llvm/PassManager.h"
19 #include "llvm/Support/CommandLine.h"
20 #include "llvm/Support/TargetRegistry.h"
21 #include "llvm/Target/TargetOptions.h"
22 #include "llvm/Transforms/Scalar.h"
26 EnableCCMP("aarch64-ccmp", cl::desc("Enable the CCMP formation pass"),
27 cl::init(true), cl::Hidden);
29 static cl::opt<bool> EnableMCR("aarch64-mcr",
30 cl::desc("Enable the machine combiner pass"),
31 cl::init(true), cl::Hidden);
34 EnableStPairSuppress("aarch64-stp-suppress", cl::desc("Suppress STP for AArch64"),
35 cl::init(true), cl::Hidden);
38 EnableAdvSIMDScalar("aarch64-simd-scalar", cl::desc("Enable use of AdvSIMD scalar"
39 " integer instructions"), cl::init(false), cl::Hidden);
42 EnablePromoteConstant("aarch64-promote-const", cl::desc("Enable the promote "
43 "constant pass"), cl::init(true), cl::Hidden);
46 EnableCollectLOH("aarch64-collect-loh", cl::desc("Enable the pass that emits the"
47 " linker optimization hints (LOH)"), cl::init(true),
51 EnableDeadRegisterElimination("aarch64-dead-def-elimination", cl::Hidden,
52 cl::desc("Enable the pass that removes dead"
53 " definitons and replaces stores to"
54 " them with stores to the zero"
59 EnableLoadStoreOpt("aarch64-load-store-opt", cl::desc("Enable the load/store pair"
60 " optimization pass"), cl::init(true), cl::Hidden);
63 EnableAtomicTidy("aarch64-atomic-cfg-tidy", cl::Hidden,
64 cl::desc("Run SimplifyCFG after expanding atomic operations"
65 " to make use of cmpxchg flow-based information"),
69 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden,
70 cl::desc("Run early if-conversion"),
74 EnableCondOpt("aarch64-condopt",
75 cl::desc("Enable the condition optimizer pass"),
76 cl::init(true), cl::Hidden);
79 EnablePBQP("aarch64-pbqp", cl::Hidden,
80 cl::desc("Use PBQP register allocator (experimental)"),
83 extern "C" void LLVMInitializeAArch64Target() {
84 // Register the target.
85 RegisterTargetMachine<AArch64leTargetMachine> X(TheAArch64leTarget);
86 RegisterTargetMachine<AArch64beTargetMachine> Y(TheAArch64beTarget);
87 RegisterTargetMachine<AArch64leTargetMachine> Z(TheARM64Target);
90 /// TargetMachine ctor - Create an AArch64 architecture model.
92 AArch64TargetMachine::AArch64TargetMachine(const Target &T, StringRef TT,
93 StringRef CPU, StringRef FS,
94 const TargetOptions &Options,
95 Reloc::Model RM, CodeModel::Model CM,
98 : LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL),
99 Subtarget(TT, CPU, FS, *this, LittleEndian), isLittle(LittleEndian),
103 if (EnablePBQP && Subtarget.isCortexA57() && OL != CodeGenOpt::None) {
105 RegisterRegAlloc::setDefault(createDefaultPBQPRegisterAllocator);
109 const AArch64Subtarget *
110 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
111 AttributeSet FnAttrs = F.getAttributes();
113 FnAttrs.getAttribute(AttributeSet::FunctionIndex, "target-cpu");
115 FnAttrs.getAttribute(AttributeSet::FunctionIndex, "target-features");
117 std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
118 ? CPUAttr.getValueAsString().str()
120 std::string FS = !FSAttr.hasAttribute(Attribute::None)
121 ? FSAttr.getValueAsString().str()
124 auto &I = SubtargetMap[CPU + FS];
126 // This needs to be done before we create a new subtarget since any
127 // creation will depend on the TM and the code generation flags on the
128 // function that reside in TargetOptions.
129 resetTargetOptions(F);
130 I = llvm::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this, isLittle);
135 void AArch64leTargetMachine::anchor() { }
137 AArch64leTargetMachine::
138 AArch64leTargetMachine(const Target &T, StringRef TT,
139 StringRef CPU, StringRef FS, const TargetOptions &Options,
140 Reloc::Model RM, CodeModel::Model CM,
141 CodeGenOpt::Level OL)
142 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
144 void AArch64beTargetMachine::anchor() { }
146 AArch64beTargetMachine::
147 AArch64beTargetMachine(const Target &T, StringRef TT,
148 StringRef CPU, StringRef FS, const TargetOptions &Options,
149 Reloc::Model RM, CodeModel::Model CM,
150 CodeGenOpt::Level OL)
151 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
154 /// AArch64 Code Generator Pass Configuration Options.
155 class AArch64PassConfig : public TargetPassConfig {
157 AArch64PassConfig(AArch64TargetMachine *TM, PassManagerBase &PM)
158 : TargetPassConfig(TM, PM) {
159 if (TM->getOptLevel() != CodeGenOpt::None)
160 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
163 AArch64TargetMachine &getAArch64TargetMachine() const {
164 return getTM<AArch64TargetMachine>();
167 void addIRPasses() override;
168 bool addPreISel() override;
169 bool addInstSelector() override;
170 bool addILPOpts() override;
171 bool addPreRegAlloc() override;
172 bool addPostRegAlloc() override;
173 bool addPreSched2() override;
174 bool addPreEmitPass() override;
178 void AArch64TargetMachine::addAnalysisPasses(PassManagerBase &PM) {
179 // Add first the target-independent BasicTTI pass, then our AArch64 pass. This
180 // allows the AArch64 pass to delegate to the target independent layer when
182 PM.add(createBasicTargetTransformInfoPass(this));
183 PM.add(createAArch64TargetTransformInfoPass(this));
186 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
187 return new AArch64PassConfig(this, PM);
190 void AArch64PassConfig::addIRPasses() {
191 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
193 addPass(createAtomicExpandPass(TM));
195 // Cmpxchg instructions are often used with a subsequent comparison to
196 // determine whether it succeeded. We can exploit existing control-flow in
197 // ldrex/strex loops to simplify this, but it needs tidying up.
198 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
199 addPass(createCFGSimplificationPass());
201 TargetPassConfig::addIRPasses();
204 // Pass Pipeline Configuration
205 bool AArch64PassConfig::addPreISel() {
206 // Run promote constant before global merge, so that the promoted constants
207 // get a chance to be merged
208 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
209 addPass(createAArch64PromoteConstantPass());
210 if (TM->getOptLevel() != CodeGenOpt::None)
211 addPass(createGlobalMergePass(TM));
212 if (TM->getOptLevel() != CodeGenOpt::None)
213 addPass(createAArch64AddressTypePromotionPass());
218 bool AArch64PassConfig::addInstSelector() {
219 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
221 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
222 // references to _TLS_MODULE_BASE_ as possible.
223 if (TM->getSubtarget<AArch64Subtarget>().isTargetELF() &&
224 getOptLevel() != CodeGenOpt::None)
225 addPass(createAArch64CleanupLocalDynamicTLSPass());
230 bool AArch64PassConfig::addILPOpts() {
232 addPass(createAArch64ConditionOptimizerPass());
234 addPass(createAArch64ConditionalCompares());
236 addPass(&MachineCombinerID);
237 if (EnableEarlyIfConversion)
238 addPass(&EarlyIfConverterID);
239 if (EnableStPairSuppress)
240 addPass(createAArch64StorePairSuppressPass());
244 bool AArch64PassConfig::addPreRegAlloc() {
245 // Use AdvSIMD scalar instructions whenever profitable.
246 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
247 addPass(createAArch64AdvSIMDScalar());
248 // The AdvSIMD pass may produce copies that can be rewritten to
249 // be register coaleascer friendly.
250 addPass(&PeepholeOptimizerID);
255 bool AArch64PassConfig::addPostRegAlloc() {
256 // Change dead register definitions to refer to the zero register.
257 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
258 addPass(createAArch64DeadRegisterDefinitions());
259 if (TM->getOptLevel() != CodeGenOpt::None &&
260 TM->getSubtarget<AArch64Subtarget>().isCortexA57() &&
261 !static_cast<const AArch64TargetMachine *>(TM)->isPBQPUsed())
262 // Improve performance for some FP/SIMD code for A57.
263 addPass(createAArch64A57FPLoadBalancing());
267 bool AArch64PassConfig::addPreSched2() {
268 // Expand some pseudo instructions to allow proper scheduling.
269 addPass(createAArch64ExpandPseudoPass());
270 // Use load/store pair instructions when possible.
271 if (TM->getOptLevel() != CodeGenOpt::None && EnableLoadStoreOpt)
272 addPass(createAArch64LoadStoreOptimizationPass());
276 bool AArch64PassConfig::addPreEmitPass() {
277 // Relax conditional branch instructions if they're otherwise out of
278 // range of their destination.
279 addPass(createAArch64BranchRelaxation());
280 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
281 TM->getSubtarget<AArch64Subtarget>().isTargetMachO())
282 addPass(createAArch64CollectLOHPass());