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 "AArch64TargetObjectFile.h"
16 #include "AArch64TargetTransformInfo.h"
17 #include "llvm/CodeGen/Passes.h"
18 #include "llvm/CodeGen/RegAllocRegistry.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/IR/LegacyPassManager.h"
21 #include "llvm/Support/CommandLine.h"
22 #include "llvm/Support/TargetRegistry.h"
23 #include "llvm/Target/TargetOptions.h"
24 #include "llvm/Transforms/Scalar.h"
28 EnableCCMP("aarch64-ccmp", cl::desc("Enable the CCMP formation pass"),
29 cl::init(true), cl::Hidden);
31 static cl::opt<bool> EnableMCR("aarch64-mcr",
32 cl::desc("Enable the machine combiner pass"),
33 cl::init(true), cl::Hidden);
36 EnableStPairSuppress("aarch64-stp-suppress", cl::desc("Suppress STP for AArch64"),
37 cl::init(true), cl::Hidden);
40 EnableAdvSIMDScalar("aarch64-simd-scalar", cl::desc("Enable use of AdvSIMD scalar"
41 " integer instructions"), cl::init(false), cl::Hidden);
44 EnablePromoteConstant("aarch64-promote-const", cl::desc("Enable the promote "
45 "constant pass"), cl::init(true), cl::Hidden);
48 EnableCollectLOH("aarch64-collect-loh", cl::desc("Enable the pass that emits the"
49 " linker optimization hints (LOH)"), cl::init(true),
53 EnableDeadRegisterElimination("aarch64-dead-def-elimination", cl::Hidden,
54 cl::desc("Enable the pass that removes dead"
55 " definitons and replaces stores to"
56 " them with stores to the zero"
61 EnableLoadStoreOpt("aarch64-load-store-opt", cl::desc("Enable the load/store pair"
62 " optimization pass"), cl::init(true), cl::Hidden);
65 EnableAtomicTidy("aarch64-atomic-cfg-tidy", cl::Hidden,
66 cl::desc("Run SimplifyCFG after expanding atomic operations"
67 " to make use of cmpxchg flow-based information"),
70 static cl::opt<bool> AArch64InterleavedAccessOpt(
71 "aarch64-interleaved-access-opt",
72 cl::desc("Optimize interleaved memory accesses in the AArch64 backend"),
73 cl::init(false), cl::Hidden);
76 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden,
77 cl::desc("Run early if-conversion"),
81 EnableCondOpt("aarch64-condopt",
82 cl::desc("Enable the condition optimizer pass"),
83 cl::init(true), cl::Hidden);
86 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden,
87 cl::desc("Work around Cortex-A53 erratum 835769"),
91 EnableGEPOpt("aarch64-gep-opt", cl::Hidden,
92 cl::desc("Enable optimizations on complex GEPs"),
95 // FIXME: Unify control over GlobalMerge.
96 static cl::opt<cl::boolOrDefault>
97 EnableGlobalMerge("aarch64-global-merge", cl::Hidden,
98 cl::desc("Enable the global merge pass"));
100 extern "C" void LLVMInitializeAArch64Target() {
101 // Register the target.
102 RegisterTargetMachine<AArch64leTargetMachine> X(TheAArch64leTarget);
103 RegisterTargetMachine<AArch64beTargetMachine> Y(TheAArch64beTarget);
104 RegisterTargetMachine<AArch64leTargetMachine> Z(TheARM64Target);
107 //===----------------------------------------------------------------------===//
108 // AArch64 Lowering public interface.
109 //===----------------------------------------------------------------------===//
110 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
111 if (TT.isOSBinFormatMachO())
112 return make_unique<AArch64_MachoTargetObjectFile>();
114 return make_unique<AArch64_ELFTargetObjectFile>();
117 // Helper function to build a DataLayout string
118 static std::string computeDataLayout(StringRef TT, bool LittleEndian) {
120 if (Triple.isOSBinFormatMachO())
121 return "e-m:o-i64:64-i128:128-n32:64-S128";
123 return "e-m:e-i64:64-i128:128-n32:64-S128";
124 return "E-m:e-i64:64-i128:128-n32:64-S128";
127 /// TargetMachine ctor - Create an AArch64 architecture model.
129 AArch64TargetMachine::AArch64TargetMachine(const Target &T, StringRef TT,
130 StringRef CPU, StringRef FS,
131 const TargetOptions &Options,
132 Reloc::Model RM, CodeModel::Model CM,
133 CodeGenOpt::Level OL,
135 // This nested ternary is horrible, but DL needs to be properly
136 // initialized before TLInfo is constructed.
137 : LLVMTargetMachine(T, computeDataLayout(TT, LittleEndian), TT, CPU, FS,
138 Options, RM, CM, OL),
139 TLOF(createTLOF(Triple(getTargetTriple()))),
140 isLittle(LittleEndian) {
144 AArch64TargetMachine::~AArch64TargetMachine() {}
146 const AArch64Subtarget *
147 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
148 Attribute CPUAttr = F.getFnAttribute("target-cpu");
149 Attribute FSAttr = F.getFnAttribute("target-features");
151 std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
152 ? CPUAttr.getValueAsString().str()
154 std::string FS = !FSAttr.hasAttribute(Attribute::None)
155 ? FSAttr.getValueAsString().str()
158 auto &I = SubtargetMap[CPU + FS];
160 // This needs to be done before we create a new subtarget since any
161 // creation will depend on the TM and the code generation flags on the
162 // function that reside in TargetOptions.
163 resetTargetOptions(F);
164 I = llvm::make_unique<AArch64Subtarget>(Triple(TargetTriple), CPU, FS,
170 void AArch64leTargetMachine::anchor() { }
172 AArch64leTargetMachine::
173 AArch64leTargetMachine(const Target &T, StringRef TT,
174 StringRef CPU, StringRef FS, const TargetOptions &Options,
175 Reloc::Model RM, CodeModel::Model CM,
176 CodeGenOpt::Level OL)
177 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
179 void AArch64beTargetMachine::anchor() { }
181 AArch64beTargetMachine::
182 AArch64beTargetMachine(const Target &T, StringRef TT,
183 StringRef CPU, StringRef FS, const TargetOptions &Options,
184 Reloc::Model RM, CodeModel::Model CM,
185 CodeGenOpt::Level OL)
186 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
189 /// AArch64 Code Generator Pass Configuration Options.
190 class AArch64PassConfig : public TargetPassConfig {
192 AArch64PassConfig(AArch64TargetMachine *TM, PassManagerBase &PM)
193 : TargetPassConfig(TM, PM) {
194 if (TM->getOptLevel() != CodeGenOpt::None)
195 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
198 AArch64TargetMachine &getAArch64TargetMachine() const {
199 return getTM<AArch64TargetMachine>();
202 void addIRPasses() override;
203 bool addPreISel() override;
204 bool addInstSelector() override;
205 bool addILPOpts() override;
206 void addPreRegAlloc() override;
207 void addPostRegAlloc() override;
208 void addPreSched2() override;
209 void addPreEmitPass() override;
213 TargetIRAnalysis AArch64TargetMachine::getTargetIRAnalysis() {
214 return TargetIRAnalysis([this](Function &F) {
215 return TargetTransformInfo(AArch64TTIImpl(this, F));
219 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
220 return new AArch64PassConfig(this, PM);
223 void AArch64PassConfig::addIRPasses() {
224 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
226 addPass(createAtomicExpandPass(TM));
228 // Cmpxchg instructions are often used with a subsequent comparison to
229 // determine whether it succeeded. We can exploit existing control-flow in
230 // ldrex/strex loops to simplify this, but it needs tidying up.
231 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
232 addPass(createCFGSimplificationPass());
234 if (TM->getOptLevel() != CodeGenOpt::None && AArch64InterleavedAccessOpt)
235 addPass(createAArch64InterleavedAccessPass());
237 TargetPassConfig::addIRPasses();
239 if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
240 // Call SeparateConstOffsetFromGEP pass to extract constants within indices
241 // and lower a GEP with multiple indices to either arithmetic operations or
242 // multiple GEPs with single index.
243 addPass(createSeparateConstOffsetFromGEPPass(TM, true));
244 // Call EarlyCSE pass to find and remove subexpressions in the lowered
246 addPass(createEarlyCSEPass());
247 // Do loop invariant code motion in case part of the lowered result is
249 addPass(createLICMPass());
253 // Pass Pipeline Configuration
254 bool AArch64PassConfig::addPreISel() {
255 // Run promote constant before global merge, so that the promoted constants
256 // get a chance to be merged
257 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
258 addPass(createAArch64PromoteConstantPass());
259 // FIXME: On AArch64, this depends on the type.
260 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
261 // and the offset has to be a multiple of the related size in bytes.
262 if ((TM->getOptLevel() != CodeGenOpt::None &&
263 EnableGlobalMerge == cl::BOU_UNSET) ||
264 EnableGlobalMerge == cl::BOU_TRUE) {
265 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
266 (EnableGlobalMerge == cl::BOU_UNSET);
267 addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize));
270 if (TM->getOptLevel() != CodeGenOpt::None)
271 addPass(createAArch64AddressTypePromotionPass());
276 bool AArch64PassConfig::addInstSelector() {
277 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
279 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
280 // references to _TLS_MODULE_BASE_ as possible.
281 if (Triple(TM->getTargetTriple()).isOSBinFormatELF() &&
282 getOptLevel() != CodeGenOpt::None)
283 addPass(createAArch64CleanupLocalDynamicTLSPass());
288 bool AArch64PassConfig::addILPOpts() {
290 addPass(createAArch64ConditionOptimizerPass());
292 addPass(createAArch64ConditionalCompares());
294 addPass(&MachineCombinerID);
295 if (EnableEarlyIfConversion)
296 addPass(&EarlyIfConverterID);
297 if (EnableStPairSuppress)
298 addPass(createAArch64StorePairSuppressPass());
302 void AArch64PassConfig::addPreRegAlloc() {
303 // Use AdvSIMD scalar instructions whenever profitable.
304 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
305 addPass(createAArch64AdvSIMDScalar());
306 // The AdvSIMD pass may produce copies that can be rewritten to
307 // be register coaleascer friendly.
308 addPass(&PeepholeOptimizerID);
312 void AArch64PassConfig::addPostRegAlloc() {
313 // Change dead register definitions to refer to the zero register.
314 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
315 addPass(createAArch64DeadRegisterDefinitions());
316 if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc())
317 // Improve performance for some FP/SIMD code for A57.
318 addPass(createAArch64A57FPLoadBalancing());
321 void AArch64PassConfig::addPreSched2() {
322 // Expand some pseudo instructions to allow proper scheduling.
323 addPass(createAArch64ExpandPseudoPass());
324 // Use load/store pair instructions when possible.
325 if (TM->getOptLevel() != CodeGenOpt::None && EnableLoadStoreOpt)
326 addPass(createAArch64LoadStoreOptimizationPass());
329 void AArch64PassConfig::addPreEmitPass() {
330 if (EnableA53Fix835769)
331 addPass(createAArch64A53Fix835769());
332 // Relax conditional branch instructions if they're otherwise out of
333 // range of their destination.
334 addPass(createAArch64BranchRelaxation());
335 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
336 Triple(TM->getTargetTriple()).isOSBinFormatMachO())
337 addPass(createAArch64CollectLOHPass());